{"schema":"5lc-index-ledger/1","key":"models-frontier-labs","layer":"models","sublayer":"frontier-labs","asOf":"2026-10-22","about":"5LC Importance Index ledger for one sublayer: the measure card, the current ten with any editorial overrides and their reasons, the computed result for every scored company (Importance = share of AI output × years to replace, Reach, replacement years, Rising, the Monte Carlo rank simulation and its stability class, validation, the revenue baseline and rater agreement), each scored company's sourced inputs and irreplaceability ratings, and the sublayer's validation cases with their sources. Monte Carlo frequencies describe the stated input ranges, not true probabilities. Omitted: the free-text notes of the internal check passes and of rating reconciliations, and the location of the internal card review; references to local evidence files and to internal method decisions are replaced by plain descriptions.","roster":{"members":[{"entityId":"anthropic","slot":"importance","since":"2026-10-22"},{"entityId":"openai","slot":"importance","since":"2026-10-22"},{"entityId":"google-deepmind","slot":"importance","since":"2026-10-22"},{"entityId":"bytedance-seed","slot":"importance","since":"2026-10-22"},{"entityId":"alibaba-qwen","slot":"importance","since":"2026-10-22"},{"entityId":"xai","slot":"importance","since":"2026-10-22"},{"entityId":"amazon-agi","slot":"importance","since":"2026-10-22"},{"entityId":"mistral","slot":"importance","since":"2026-10-22"},{"entityId":"baidu-ernie","slot":"importance","since":"2026-10-22"},{"entityId":"meta-superintelligence-labs","slot":"importance","since":"2026-10-22"}],"pending":{},"lastRun":"2026-10-22"},"card":{"schema":"5lc-index-card/1","key":"models-frontier-labs","layer":"models","sublayer":"frontier-labs","version":1,"scope":"Paid API access to closed general-purpose and coding models (models whose weights are not published), including their embedding and real-time voice versions, counted for the model's developer whether bought from the developer, through a cloud platform such as Amazon Bedrock, Google Vertex AI or Microsoft Foundry, or as a licence to another platform. Consumer apps and subscriptions count in Applications · Consumer, business seats in Enterprise agents, coding-tool plans in Developer tools, open-weight models in Open-weight, serving by companies that did not build the model in Inference serving, and image, video, music and speech-only models in Specialist models.","outputMeasure":"API spending on closed general-purpose and coding models, counted for each model's developer.","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","hardToReplace":"Customers can switch models within weeks through an API. Top-level capability is hard to replace, above all in coding agents, where one or two developers lead; a new frontier model takes years and very large computing budgets.","functions":[{"id":"coding-agents","name":"Coding and software agents","why":"API use of general and code-specialised models for writing and running code, the largest single workload; a few models lead it.","weight":{"low":0.255,"central":0.348,"high":0.545},"weightSource":{"url":"https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/","title":"2025: The State of Generative AI in the Enterprise","publisher":"Menlo Ventures","publishedDate":"2025-12-09","accessedDate":"2026-10-22","tier":"T4","locator":"Enterprise LLM API share of spend: Anthropic 40%, OpenAI 27%, Google 21%; coding share: Anthropic 54%, OpenAI 21% (survey of 495 US enterprise AI decision-makers, November 2025)"},"shareData":"Coding spend shares from Menlo Ventures surveys and gateway data, scaled to each developer's reported API revenue; units of larger groups from spend-share or usage data, never the parent's revenue"},{"id":"general-reasoning","name":"General reasoning, text and agents","why":"Chat, analysis, search, customer service and other agents: most of the spending, with more alternatives.","weight":{"low":0.413,"central":0.631,"high":0.735},"weightSource":{"url":"https://www.anthropic.com/research/economic-index-primitives","title":"Anthropic Economic Index: New building blocks for understanding AI use","publisher":"Anthropic","publishedDate":"2026-01-15","accessedDate":"2026-10-22","tier":"T2","locator":"November 2025 sample of 1 million first-party API transcripts: computer and mathematical tasks are nearly half of API traffic; other occupations make up the rest"},"shareData":"Enterprise and gateway spend shares net of coding, and reported API revenue; usage share for units of larger groups"},{"id":"china-api","name":"Closed models in mainland China","why":"A separate market without the main foreign models; small in revenue at Chinese prices.","weight":{"low":0.009,"central":0.021,"high":0.042},"weightSource":{"url":"https://www.idc.com/?p=114742","title":"中国MaaS市场进入高速增长期，Token经济从概念走向规模化实践 (China's MaaS market enters a high-growth phase)","publisher":"IDC","publishedDate":"2026-05-07","accessedDate":"2026-10-22","tier":"T4","locator":"China public-cloud MaaS revenue RMB 3.07bn in 2025, 2026 forecast about RMB 18.6bn; Volcano Engine above 40% of revenue, followed by Alibaba Cloud, Baidu AI Cloud, Zhipu and China Mobile Cloud (IDC China AI Software Market Half-Year Tracker, 2025H2)"},"shareData":"IDC platform shares, closed models only, and company disclosures"}],"risingFloorPct":0.25,"validationCaseIds":[],"knownGaps":["No public source splits model API spending by workload; the coding weight combines Menlo Ventures' coding shares with Anthropic's own API traffic mix (November 2025 data), before most of 2026's growth in coding agents. Menlo Ventures is an investor in Anthropic.","Developers count sales through cloud partners differently: Anthropic includes them in revenue and OpenAI does not, so shares must be rebuilt on the card's basis.","Sales through cloud platforms are counted in full for the model's developer here, and the platform's retained part is counted again in Inference serving.","OpenAI, Google and SpaceX (xAI) do not report API revenue separately from subscriptions; SpaceX's Super Intelligence segment also includes Grok subscriptions and Cursor. Their shares come from spend-share and usage data.","Spending on OpenAI models sold through Microsoft Azure is not disclosed; the weights use a 5LayerCake estimate (US$3–12bn a year) built from reported revenue-share payments.","IDC's China totals mix closed models with open models served on the same platforms; the closed-model portion is estimated (40 to 70%).","Spend-share data come from samples (US enterprises, developer gateways), not the whole market."],"approval":{"approver":"Claude (delegated by Dachi)","date":"2026-10-22","sha256":"f8854b58913974d282ce032b3e88032b7984c21b72c1dd73d92fb9a1a8265c84","delegatedOn":"2026-10-22"},"risingFloorReason":"One developer holds about three-fifths of closed-model API spending and only four or five companies reach 1%, all within the eight Importance places; 0.25% lets fast-growing smaller developers compete for the two Rising places, as approved for Chip design and Foundry."},"result":{"schema":"5lc-index-result/1","key":"models-frontier-labs","asOf":"2026-10-22","methodVersion":"1.1","methodSha256":"3cabed6b38514c263d8b86a494e8cc19cdbd0b2a1013543dc024192c49f7a6a6","cardSha256":"f8854b58913974d282ce032b3e88032b7984c21b72c1dd73d92fb9a1a8265c84","validationSha256":"409931bb458f38b02342c9bf43e61eab299e47d8679912af2c96074ec556f17e","inputsSha256":"4af075c9c8cf2d7b3b10dc6c5f9d45f03fbb011a46fd066bd362f2a466ed7806","unscreened":[],"companies":[{"entityId":"anthropic","name":"Anthropic","reachPct":63.4353,"importance":37.2889052566,"importanceRank":1,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":77.3,"shareLowPct":70.2,"shareHighPct":82.3,"stale":false,"offBasis":false,"level":2.5,"years":0.7071067811865475,"disagreement":false},{"id":"general-reasoning","weight":0.631,"sharePct":57.9,"shareLowPct":46.9,"shareHighPct":63.4,"stale":false,"offBasis":false,"level":2,"years":0.5,"disagreement":false},{"id":"china-api","weight":0.021,"sharePct":0,"shareLowPct":0,"shareHighPct":0,"stale":false,"offBasis":false,"level":null,"years":null,"disagreement":false}],"rising":91.6666666667,"risingEligible":true,"risingComponents":{"growth":83.3333333333,"shareChange":100},"mc":{"medianRank":1,"rankP10":1,"rankP90":1,"pTop8":1,"pSelected":1,"stability":"firm"},"baselineRevenueRank":1},{"entityId":"openai","name":"OpenAI","reachPct":20.003,"importance":7.55072147219,"importanceRank":2,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":13.6,"shareLowPct":5.66,"shareHighPct":26.5,"stale":false,"offBasis":false,"level":1.33333333333,"years":0.3149802624729905,"disagreement":false},{"id":"general-reasoning","weight":0.631,"sharePct":24.2,"shareLowPct":13.2,"shareHighPct":38.7,"stale":false,"offBasis":false,"level":1.66666666667,"years":0.3968502629929668,"disagreement":false},{"id":"china-api","weight":0.021,"sharePct":0,"shareLowPct":0,"shareHighPct":0,"stale":false,"offBasis":false,"level":null,"years":null,"disagreement":false}],"rising":0,"risingEligible":true,"risingComponents":{"growth":0,"shareChange":0},"mc":{"medianRank":2,"rankP10":2,"rankP90":2,"pTop8":1,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"google-deepmind","name":"Google DeepMind","reachPct":9.81312,"importance":2.75371369388,"importanceRank":3,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":6.44,"shareLowPct":2.24,"shareHighPct":13.6,"stale":false,"offBasis":false,"level":1.16666666667,"years":0.2806155120779916,"disagreement":false},{"id":"general-reasoning","weight":0.631,"sharePct":12,"shareLowPct":5.91,"shareHighPct":22.4,"stale":false,"offBasis":false,"level":1.16666666667,"years":0.2806155120779916,"disagreement":false},{"id":"china-api","weight":0.021,"sharePct":0,"shareLowPct":0,"shareHighPct":0,"stale":false,"offBasis":false,"level":null,"years":null,"disagreement":false}],"rising":41.6666666667,"risingEligible":true,"risingComponents":{"growth":33.3333333333,"shareChange":50},"mc":{"medianRank":3,"rankP10":3,"rankP90":3,"pTop8":1,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"bytedance-seed","name":"ByteDance Seed","reachPct":1.0417752,"importance":0.395516671901,"importanceRank":4,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.0894,"shareLowPct":0.00895,"shareHighPct":0.593,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.144,"shareLowPct":0.0207,"shareHighPct":0.697,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":43.8,"shareLowPct":21.8,"shareHighPct":57.5,"stale":false,"offBasis":false,"level":1.66666666667,"years":0.3968502629929668,"disagreement":false}],"rising":null,"risingEligible":true,"risingComponents":{"growth":83.3333333333},"mc":{"medianRank":4,"rankP10":4,"rankP90":6,"pTop8":0.9999,"pSelected":1,"stability":"firm"},"baselineRevenueRank":2},{"entityId":"alibaba-qwen","name":"Alibaba Qwen","reachPct":0.7495282,"importance":0.253707996675,"importanceRank":5,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.0939,"shareLowPct":0.0119,"shareHighPct":0.446,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.121,"shareLowPct":0.0206,"shareHighPct":0.493,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":30.5,"shareLowPct":15.1,"shareHighPct":54.1,"stale":false,"offBasis":false,"level":1.5,"years":0.3535533905932738,"disagreement":false}],"rising":null,"risingEligible":true,"risingComponents":{},"mc":{"medianRank":6,"rankP10":5,"rankP90":7,"pTop8":0.9984,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"xai","name":"xAI (SpaceXAI)","reachPct":0.9813008,"importance":0.2453252,"importanceRank":6,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.715,"shareLowPct":0.135,"shareHighPct":2.92,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":1.16,"shareLowPct":0.312,"shareHighPct":3.43,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":true,"risingComponents":{},"mc":{"medianRank":5,"rankP10":4,"rankP90":7,"pTop8":0.9983,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"amazon-agi","name":"Amazon AGI (Nova)","reachPct":0.5051438,"importance":0.12628595,"importanceRank":7,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.179,"shareLowPct":0.0149,"shareHighPct":1.06,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.701,"shareLowPct":0.145,"shareHighPct":2.31,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":true,"risingComponents":{},"mc":{"medianRank":7,"rankP10":5,"rankP90":8,"pTop8":0.9508,"pSelected":0.9929,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"mistral","name":"Mistral AI","reachPct":0.2473518,"importance":0.06183795,"importanceRank":8,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.149,"shareLowPct":0.0119,"shareHighPct":0.734,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.309,"shareLowPct":0.0455,"shareHighPct":1.01,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":8,"rankP10":7,"rankP90":11,"pTop8":0.5343,"pSelected":0.8606,"stability":"likely"},"baselineRevenueRank":null},{"entityId":"baidu-ernie","name":"Baidu ERNIE","reachPct":0.16851558,"importance":0.042128895,"importanceRank":9,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00596,"shareLowPct":0,"shareHighPct":0.0596,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0165,"shareLowPct":0,"shareHighPct":0.0979,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":7.43,"shareLowPct":2.61,"shareHighPct":15.1,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":12,"rankP10":9,"rankP90":15,"pTop8":0.0433,"pSelected":0.314,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"meta-superintelligence-labs","name":"Meta Superintelligence Labs","reachPct":0.146399,"importance":0.03659975,"importanceRank":10,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.107,"shareLowPct":0.00895,"shareHighPct":0.741,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.173,"shareLowPct":0.0207,"shareHighPct":0.87,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0,"shareLowPct":0,"shareHighPct":0,"stale":false,"offBasis":false,"level":null,"years":null,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":9,"rankP10":8,"rankP90":13,"pTop8":0.3141,"pSelected":0.7174,"stability":"likely"},"baselineRevenueRank":null},{"entityId":"tencent-hy","name":"Tencent Hy","reachPct":0.1067637,"importance":0.026690925,"importanceRank":11,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.0179,"shareLowPct":0.000745,"shareHighPct":0.143,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0495,"shareLowPct":0.00619,"shareHighPct":0.235,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":3.3,"shareLowPct":0.847,"shareHighPct":9.29,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":13,"rankP10":10,"rankP90":17,"pTop8":0.0074,"pSelected":0.1311,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"cohere","name":"Cohere","reachPct":0.10440276,"importance":0.02610069,"importanceRank":12,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00477,"shareLowPct":0,"shareHighPct":0.0745,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.162,"shareLowPct":0.0372,"shareHighPct":0.514,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":14,"rankP10":10,"rankP90":19,"pTop8":0.0093,"pSelected":0.1282,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"iflytek","name":"iFlytek (Spark)","reachPct":0.0756742,"importance":0.01891855,"importanceRank":13,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00149,"shareLowPct":0,"shareHighPct":0.0268,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.00928,"shareLowPct":0,"shareHighPct":0.0618,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":3.3,"shareLowPct":0.425,"shareHighPct":20.4,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":13,"rankP10":9,"rankP90":19,"pTop8":0.0688,"pSelected":0.2749,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"typesafe-ai","name":"TypeSafe AI","reachPct":0.0649514,"importance":0.01623785,"importanceRank":14,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00745,"shareLowPct":0,"shareHighPct":0.208,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.098,"shareLowPct":0.00826,"shareHighPct":0.717,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":13,"rankP10":9,"rankP90":19,"pTop8":0.0593,"pSelected":0.3044,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"sber","name":"Sber (GigaChat)","reachPct":0.0642628,"importance":0.0160657,"importanceRank":15,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.0149,"shareLowPct":0.000596,"shareHighPct":0.186,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0928,"shareLowPct":0.0155,"shareHighPct":0.503,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":15,"rankP10":10,"rankP90":20,"pTop8":0.0098,"pSelected":0.1195,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"perplexity","name":"Perplexity (Sonar API)","reachPct":0.05239694,"importance":0.013099235,"importanceRank":16,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00238,"shareLowPct":0,"shareHighPct":0.0358,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0809,"shareLowPct":0.0186,"shareHighPct":0.247,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":20,"rankP10":16,"rankP90":24,"pTop8":0,"pSelected":0.0006,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"naver-hyperclova","name":"NAVER HyperCLOVA X","reachPct":0.05206528,"importance":0.01301632,"importanceRank":17,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00596,"shareLowPct":0,"shareHighPct":0.0894,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0784,"shareLowPct":0.0165,"shareHighPct":0.309,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":18,"rankP10":14,"rankP90":23,"pTop8":0.0001,"pSelected":0.008,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"moonshot","name":"Moonshot AI","reachPct":0.0499712,"importance":0.0124928,"importanceRank":18,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.0238,"shareLowPct":0,"shareHighPct":0.268,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.247,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":1.24,"shareLowPct":0,"shareHighPct":7.71,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":14,"rankP10":11,"rankP90":18,"pTop8":0.0042,"pSelected":0.0973,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"sensetime","name":"SenseTime (SenseNova)","reachPct":0.0496342,"importance":0.01240855,"importanceRank":19,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00149,"shareLowPct":0,"shareHighPct":0.0268,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.00928,"shareLowPct":0,"shareHighPct":0.0618,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":2.06,"shareLowPct":0.42,"shareHighPct":6.32,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":20,"rankP10":16,"rankP90":23,"pTop8":0,"pSelected":0.0003,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"microsoft-ai","name":"Microsoft AI (MAI)","reachPct":0.0363893,"importance":0.009097325,"importanceRank":20,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.0358,"shareLowPct":0.00298,"shareHighPct":0.357,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0371,"shareLowPct":0.00413,"shareHighPct":0.329,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":16,"rankP10":11,"rankP90":21,"pTop8":0.0019,"pSelected":0.0496,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"china-mobile","name":"China Mobile (Jiutian)","reachPct":0.035927364,"importance":0.008981841,"importanceRank":21,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.000298,"shareLowPct":0,"shareHighPct":0.00894,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.00186,"shareLowPct":0,"shareHighPct":0.0206,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":1.65,"shareLowPct":0.418,"shareHighPct":4.8,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":22,"rankP10":19,"rankP90":26,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"zhipu","name":"Zhipu AI (Z.ai)","reachPct":0.03369997,"importance":0.0084249925,"importanceRank":22,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00596,"shareLowPct":0,"shareHighPct":0.0894,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.00619,"shareLowPct":0,"shareHighPct":0.0824,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":1.32,"shareLowPct":0,"shareHighPct":6.27,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":19,"rankP10":15,"rankP90":23,"pTop8":0.0001,"pSelected":0.0012,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"huawei-cloud","name":"Huawei Cloud (Pangu)","reachPct":0.0324142,"importance":0.00810355,"importanceRank":23,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00149,"shareLowPct":0,"shareHighPct":0.0268,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.00928,"shareLowPct":0,"shareHighPct":0.0618,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":1.24,"shareLowPct":0.25,"shareHighPct":4.78,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":21,"rankP10":18,"rankP90":25,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"stepfun","name":"StepFun","reachPct":0.01632751,"importance":0.0040818775,"importanceRank":24,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00179,"shareLowPct":0,"shareHighPct":0.0298,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.00289,"shareLowPct":0,"shareHighPct":0.0371,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.661,"shareLowPct":0,"shareHighPct":3.22,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":24,"rankP10":20,"rankP90":28,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"kuaishou-kwaipilot","name":"Kuaishou Kwaipilot (KAT-Coder)","reachPct":0.01398613,"importance":0.0034965325,"importanceRank":25,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.0134,"shareLowPct":0.00209,"shareHighPct":0.0894,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.00103,"shareLowPct":0,"shareHighPct":0.0186,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.413,"shareLowPct":0.0828,"shareHighPct":1.63,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":25,"rankP10":22,"rankP90":28,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"writer","name":"Writer","reachPct":0.013474408,"importance":0.003368602,"importanceRank":26,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.000596,"shareLowPct":0,"shareHighPct":0.0119,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0202,"shareLowPct":0.00371,"shareHighPct":0.0825,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":28,"rankP10":24,"rankP90":29,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"ntt","name":"NTT (tsuzumi)","reachPct":0.01340692,"importance":0.00335173,"importanceRank":27,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00149,"shareLowPct":0,"shareHighPct":0.0179,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0196,"shareLowPct":0,"shareHighPct":0.0619,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":28,"rankP10":26,"rankP90":29,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"upstage","name":"Upstage","reachPct":0.01340692,"importance":0.00335173,"importanceRank":28,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00149,"shareLowPct":0,"shareHighPct":0.0298,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.0196,"shareLowPct":0.00495,"shareHighPct":0.103,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":26,"rankP10":22,"rankP90":29,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"minimax","name":"MiniMax","reachPct":0.01118797,"importance":0.0027969925,"importanceRank":29,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00596,"shareLowPct":0,"shareHighPct":0.0536,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.00619,"shareLowPct":0,"shareHighPct":0.0495,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.248,"shareLowPct":0,"shareHighPct":1.63,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":25,"rankP10":22,"rankP90":28,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"inception-labs","name":"Inception","reachPct":0.00410765,"importance":0.0010269125,"importanceRank":30,"functions":[{"id":"coding-agents","weight":0.348,"sharePct":0.00358,"shareLowPct":0.000298,"shareHighPct":0.0179,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"general-reasoning","weight":0.631,"sharePct":0.00371,"shareLowPct":0.000413,"shareHighPct":0.0165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"china-api","weight":0.021,"sharePct":0.0248,"shareLowPct":0,"shareHighPct":0.165,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":30,"rankP10":29,"rankP90":30,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null}],"selection":{"importanceIds":["anthropic","openai","google-deepmind","bytedance-seed","alibaba-qwen","xai","amazon-agi","mistral"],"risingIds":[],"filledFromImportance":["baidu-ernie","meta-superintelligence-labs"]},"risingPools":{"growth":4,"shareChange":3,"committed":0},"weightSensitivity":[{"functionId":"coding-agents","end":"low","entered":[],"left":[]},{"functionId":"coding-agents","end":"high","entered":[],"left":[]},{"functionId":"general-reasoning","end":"low","entered":["tencent-hy"],"left":["meta-superintelligence-labs"]},{"functionId":"general-reasoning","end":"high","entered":[],"left":[]},{"functionId":"china-api","end":"low","entered":["cohere"],"left":["baidu-ernie"]},{"functionId":"china-api","end":"high","entered":["tencent-hy"],"left":["meta-superintelligence-labs"]}],"validation":[],"baseline":{"spearman":1,"baselineTop10NotSelected":[],"selectedNotInBaselineTop10":["openai","google-deepmind","alibaba-qwen","xai","amazon-agi","mistral","baidu-ernie","meta-superintelligence-labs"]},"raterAgreement":{"within":1,"weightedKappa":0.779220779221,"disagreements":[]}},"inputs":{"anthropic":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"anthropic","name":"Anthropic","kind":"private","commercialActivityVerified":true,"shares":{"coding-agents":{"low":70.2,"central":77.3,"high":82.3,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://thenextweb.com/news/anthropic-q3-revenue-16-8bn-ipo","title":"Anthropic tells investors Q3 revenue reached $16.8bn, FT reports","publisher":"The Next Web","publishedDate":"2026-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"Q3 2026 revenue $16.8bn (Q2 about $11.6bn, Q1 about $4.7bn); first reported by the FT, then Reuters and Bloomberg; Anthropic has not published the figures"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Anthropic: Anthropic told investors that Q3 2026 revenue was $16.8bn (FT first, then Reuters and Bloomberg; T3): × 4 = $67.2bn annualised, which includes sales through cloud partners (Anthropic books them gross; Reuters, 30 Sep 2026, https://www.khaleejtimes.com/business/anthropic-ipo-prospectus-lays-bare-deep-dependence-on-big-tech-partners, opened). API share of that revenue: Reuters' report on Anthropic's IPO filing gives 2025 usage-based revenue of about $3.8bn against $789m of subscriptions, 83% usage-based (T3), and says Anthropic expects usage-based revenue to remain 'the substantial majority'. 5LayerCake assumes 80% (72-86%) for Q3 2026, a little below 2025 to allow for faster growth of subscription plans (assumption). 67.2 × 0.80 = US$53.8bn, rounded to $54bn (45-59; low 63 × 0.72 = 45.4, high 69 × 0.86 = 59.3). Coding fraction 0.48 (0.40-0.60): Anthropic says computer and mathematical tasks are 'nearly half' of its first-party API traffic (Economic Index, November 2025 sample of 1m transcripts, T2, https://www.anthropic.com/research/economic-index-primitives, opened); Menlo Ventures (9 Dec 2025, T4, https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/, opened) gives Anthropic 54% of coding spend against 40% of all enterprise LLM API spend, consistent with a coding fraction near one half (Menlo is an Anthropic investor). Cross-check, not used as an input: Vercel's gateway had Anthropic at 43.9% of September 2026 spend, down from 63.7% in August (T2), in a developer sample where open-weight models were nearly a quarter of spend. Coding spend = 54 × 0.48 = US$25.9bn (low 45 × 0.4 = 18bn; high 59 × 0.6 = 35.4bn). Part total US$33.54bn. Share 77.3% (range 70.2-82.3%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-10-15 (full date)."},"general-reasoning":{"low":46.9,"central":57.9,"high":63.4,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://thenextweb.com/news/anthropic-q3-revenue-16-8bn-ipo","title":"Anthropic tells investors Q3 revenue reached $16.8bn, FT reports","publisher":"The Next Web","publishedDate":"2026-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"Q3 2026 revenue $16.8bn (Q2 about $11.6bn, Q1 about $4.7bn); first reported by the FT, then Reuters and Bloomberg; Anthropic has not published the figures"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Anthropic: Anthropic told investors that Q3 2026 revenue was $16.8bn (FT first, then Reuters and Bloomberg; T3): × 4 = $67.2bn annualised, which includes sales through cloud partners (Anthropic books them gross; Reuters, 30 Sep 2026, https://www.khaleejtimes.com/business/anthropic-ipo-prospectus-lays-bare-deep-dependence-on-big-tech-partners, opened). API share of that revenue: Reuters' report on Anthropic's IPO filing gives 2025 usage-based revenue of about $3.8bn against $789m of subscriptions, 83% usage-based (T3), and says Anthropic expects usage-based revenue to remain 'the substantial majority'. 5LayerCake assumes 80% (72-86%) for Q3 2026, a little below 2025 to allow for faster growth of subscription plans (assumption). 67.2 × 0.80 = US$53.8bn, rounded to $54bn (45-59; low 63 × 0.72 = 45.4, high 69 × 0.86 = 59.3). Coding fraction 0.48 (0.40-0.60): Anthropic says computer and mathematical tasks are 'nearly half' of its first-party API traffic (Economic Index, November 2025 sample of 1m transcripts, T2, https://www.anthropic.com/research/economic-index-primitives, opened); Menlo Ventures (9 Dec 2025, T4, https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/, opened) gives Anthropic 54% of coding spend against 40% of all enterprise LLM API spend, consistent with a coding fraction near one half (Menlo is an Anthropic investor). Cross-check, not used as an input: Vercel's gateway had Anthropic at 43.9% of September 2026 spend, down from 63.7% in August (T2), in a developer sample where open-weight models were nearly a quarter of spend. General spend = 54 × (1 − 0.48) = US$28.1bn (low 45 × 0.4 = 18bn; high 59 × 0.6 = 35.4bn). Part total US$48.48bn. Share 57.9% (range 46.9-63.4%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-10-15 (full date)."},"china-api":{"participates":false,"evidence":{"url":"https://www.anthropic.com/supported-countries","title":"Supported countries and regions","publisher":"Anthropic","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T2","locator":"Mainland China is not on the list; 'any country or region not listed is unsupported'; the page shows a policy effective date of 12 Nov 2026 and no publication date"}}},"screening":{"coding-agents":{"levelLow":2,"levelHigh":4},"general-reasoning":{"levelLow":2,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c75fa11e7e5bbd81bbc76feb4da51cae9a518af2ce937fa75a5663be75c16445"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c75fa11e7e5bbd81bbc76feb4da51cae9a518af2ce937fa75a5663be75c16445"}},"activityGrowth":{"current":{"low":67200000000,"central":67200000000,"high":67200000000,"unit":"USD","basis":"Annualised revenue run-rate, all products and channels (latest month × 12 or latest quarter × 4), US dollars","period":"Q3 2026 revenue × 4","periodEnd":"2026-09-30","source":{"url":"https://thenextweb.com/news/anthropic-q3-revenue-16-8bn-ipo","title":"Anthropic tells investors Q3 revenue reached $16.8bn, FT reports","publisher":"The Next Web","publishedDate":"2026-10-15","accessedDate":"2026-10-22","tier":"T3","locator":"Q3 2026 revenue $16.8bn (Q2 about $11.6bn, Q1 about $4.7bn); first reported by the FT, then Reuters and Bloomberg; Anthropic has not published the figures"},"note":"Anthropic told investors on 14 Oct 2026 that Q3 2026 revenue was $16.8bn (FT, Reuters and Bloomberg); × 4 = $67.2bn. Total revenue is used as the measure of API activity: usage-based billing was 83% of 2025 revenue and is expected to stay the substantial majority (Reuters, 30 Sep 2026)."},"prior":{"low":7000000000,"central":7000000000,"high":7000000000,"unit":"USD","basis":"Annualised revenue run-rate, all products and channels (latest month × 12 or latest quarter × 4), US dollars","period":"October 2025 run-rate","periodEnd":"2025-10-31","source":{"url":"https://the-decoder.com/anthropic-aims-to-triple-annual-revenue-to-over-20-billion-by-2026/","title":"Anthropic aims to triple annual revenue to over $20 billion by 2026","publisher":"The Decoder (reporting Reuters)","publishedDate":"2025-10-16","accessedDate":"2026-10-22","tier":"T2","locator":"Anthropic told Reuters its annualised revenue was around $7bn as of October 2025; $9bn year-end target; enterprise clients about 80% of revenue"},"note":"Anthropic told Reuters its annualised revenue was around $7bn as of October 2025 (company statement, reported by Reuters on 16 Oct 2025). Growth is about 9.6 times, above the engine's +300% cap."}},"baselineRevenueUsd":{"low":3800000000,"central":3800000000,"high":3800000000,"unit":"USD","basis":"Usage-based (consumption) revenue, calendar year, US dollars (includes sales through cloud marketplaces)","period":"CY2025","periodEnd":"2025-12-31","source":{"url":"https://www.khaleejtimes.com/business/anthropic-ipo-prospectus-lays-bare-deep-dependence-on-big-tech-partners","title":"Anthropic IPO prospectus lays bare deep dependence on Big Tech partners","publisher":"Reuters (via Khaleej Times)","publishedDate":"2026-09-30","accessedDate":"2026-10-22","tier":"T3","locator":"2025 usage-based revenue about $3.8bn, subscriptions $789m, total nearly $4.6bn; cloud marketplaces $2.16bn (47%); 'the substantial majority' of future revenue expected to be usage-based"},"note":"Reuters, citing Anthropic's IPO filing: about $3.8bn of 2025 revenue came from customers paying by usage, against $789m of subscriptions."},"priorShares":{"coding-agents":{"low":52.2,"central":56.6,"high":59.7,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2025, annualised","periodEnd":"2025-09-30","source":{"url":"https://the-decoder.com/anthropic-aims-to-triple-annual-revenue-to-over-20-billion-by-2026/","title":"Anthropic aims to triple annual revenue to over $20 billion by 2026","publisher":"The Decoder (reporting Reuters)","publishedDate":"2025-10-16","accessedDate":"2026-10-22","tier":"T5","locator":"Anthropic told Reuters its annualised revenue was around $7bn as of October 2025; $9bn year-end target; enterprise clients about 80% of revenue"},"note":"5LayerCake estimate (T5). Prior share: Anthropic's annualised revenue was around $7bn in October 2025 (company statement via Reuters, T2) against $67.2bn now, with the API share taken as 83% (Reuters, FY2025 usage-based share, T3) in 2025: US$5.8bn (4.9-6.6). The market total for Q3 2025 is a frame built the same way as the current one: OpenAI's spend and Google's spend scaled by their own activity growth (see their notes) and every other company scaled down by the market growth these three imply (6.02 times), so no change in share is claimed for them. Coding fractions are assumed unchanged. Spend in the part = 5.81 × 0.48 = US$2.79bn; part total US$4.93bn (central). Share 56.6% (range 52.2-59.7%). The current share (77.3%) is on the same basis, so the change is +20.7 points."},"general-reasoning":{"low":30.9,"central":34.7,"high":37.7,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2025, annualised","periodEnd":"2025-09-30","source":{"url":"https://the-decoder.com/anthropic-aims-to-triple-annual-revenue-to-over-20-billion-by-2026/","title":"Anthropic aims to triple annual revenue to over $20 billion by 2026","publisher":"The Decoder (reporting Reuters)","publishedDate":"2025-10-16","accessedDate":"2026-10-22","tier":"T5","locator":"Anthropic told Reuters its annualised revenue was around $7bn as of October 2025; $9bn year-end target; enterprise clients about 80% of revenue"},"note":"5LayerCake estimate (T5). Prior share: Anthropic's annualised revenue was around $7bn in October 2025 (company statement via Reuters, T2) against $67.2bn now, with the API share taken as 83% (Reuters, FY2025 usage-based share, T3) in 2025: US$5.8bn (4.9-6.6). The market total for Q3 2025 is a frame built the same way as the current one: OpenAI's spend and Google's spend scaled by their own activity growth (see their notes) and every other company scaled down by the market growth these three imply (6.02 times), so no change in share is claimed for them. Coding fractions are assumed unchanged. Spend in the part = 5.81 × 0.52 = US$3.02bn; part total US$8.702bn (central). Share 34.7% (range 30.9-37.7%). The current share (57.9%) is on the same basis, so the change is +23.2 points."}}},"openai":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"openai","name":"OpenAI","kind":"private","commercialActivityVerified":true,"shares":{"coding-agents":{"low":5.66,"central":13.6,"high":26.5,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.zawya.com/en/special-coverage/openais-september-annualized-revenue-nears-50bln-less-than-previously-indicated-source-says-1586243","title":"OpenAI's September annualized revenue nears $50bln, less than previously indicated, source says","publisher":"Reuters (via Zawya)","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Annualised revenue 'almost $50 billion' in September; excludes sales through cloud partners such as AWS and Google Cloud, which Anthropic includes; the page shows no date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for OpenAI: OpenAI's annualised revenue was 'almost $50bn' at the end of September 2026 (Reuters via Zawya, T3; taken as 49), excluding sales through cloud partners such as AWS and Google Cloud. OpenAI does not report an API share. 5LayerCake assumes 20% (12-30%) of that revenue: in October 2024 about 75% of revenue was consumer subscriptions and the rest API and enterprise, and in August 2026 its CFO reportedly said enterprise now exceeds consumer (Tech Times, 15 Aug 2026, https://www.techtimes.com/articles/324562/20260815/openai-enterprise-revenue-tops-consumer-first-time-40-billion-arr-two-quarters-early.htm, opened; investor meeting reported by Bloomberg and CNBC), with no API split given. 49 × 0.20 = US$9.8bn of its own API. Plus OpenAI models sold through Microsoft Azure: US$6bn (3-12), the card's 5LayerCake estimate built from reported revenue-share payments (TechCrunch, 14 Nov 2025, not re-opened); Microsoft's FY2026 10-K shows US$24.1bn of revenue from commercial arrangements with OpenAI including revenue-share payments, which is mostly compute sold to OpenAI and cannot be split to give Azure-sold model spending, so it is not used. Plus OpenAI models sold through Amazon Bedrock since 2026 (GPT-5.4 and later added; Amazon Q1 and Q2 2026 calls, https://www.fool.com/earnings/call-transcripts/2026/08/07/amazon-amzn-q2-2026-earnings-call-transcript/, opened): assumed US$0.5bn (0.1-2). Total US$16.3bn (low 47 × 0.12 + 3 + 0.1 = 8.7; high 50 × 0.30 + 12 + 2 = 29). Coding fraction 0.28 (0.20-0.36): OpenAI has 21% of coding spend against 27% of all spend (Menlo, T4, opened), which with Anthropic's 54% and 40% gives a coding fraction of 0.576 × Anthropic's, 0.28 at 0.48 (0.23-0.35 over Anthropic's range), widened for survey error. Coding spend = 16.3 × 0.28 = US$4.56bn (low 8.7 × 0.2 = 1.74bn; high 29 × 0.36 = 10.4bn). Part total US$33.54bn. Share 13.6% (range 5.66-26.5%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: undated (the page shows none)."},"general-reasoning":{"low":13.2,"central":24.2,"high":38.7,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.zawya.com/en/special-coverage/openais-september-annualized-revenue-nears-50bln-less-than-previously-indicated-source-says-1586243","title":"OpenAI's September annualized revenue nears $50bln, less than previously indicated, source says","publisher":"Reuters (via Zawya)","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Annualised revenue 'almost $50 billion' in September; excludes sales through cloud partners such as AWS and Google Cloud, which Anthropic includes; the page shows no date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for OpenAI: OpenAI's annualised revenue was 'almost $50bn' at the end of September 2026 (Reuters via Zawya, T3; taken as 49), excluding sales through cloud partners such as AWS and Google Cloud. OpenAI does not report an API share. 5LayerCake assumes 20% (12-30%) of that revenue: in October 2024 about 75% of revenue was consumer subscriptions and the rest API and enterprise, and in August 2026 its CFO reportedly said enterprise now exceeds consumer (Tech Times, 15 Aug 2026, https://www.techtimes.com/articles/324562/20260815/openai-enterprise-revenue-tops-consumer-first-time-40-billion-arr-two-quarters-early.htm, opened; investor meeting reported by Bloomberg and CNBC), with no API split given. 49 × 0.20 = US$9.8bn of its own API. Plus OpenAI models sold through Microsoft Azure: US$6bn (3-12), the card's 5LayerCake estimate built from reported revenue-share payments (TechCrunch, 14 Nov 2025, not re-opened); Microsoft's FY2026 10-K shows US$24.1bn of revenue from commercial arrangements with OpenAI including revenue-share payments, which is mostly compute sold to OpenAI and cannot be split to give Azure-sold model spending, so it is not used. Plus OpenAI models sold through Amazon Bedrock since 2026 (GPT-5.4 and later added; Amazon Q1 and Q2 2026 calls, https://www.fool.com/earnings/call-transcripts/2026/08/07/amazon-amzn-q2-2026-earnings-call-transcript/, opened): assumed US$0.5bn (0.1-2). Total US$16.3bn (low 47 × 0.12 + 3 + 0.1 = 8.7; high 50 × 0.30 + 12 + 2 = 29). Coding fraction 0.28 (0.20-0.36): OpenAI has 21% of coding spend against 27% of all spend (Menlo, T4, opened), which with Anthropic's 54% and 40% gives a coding fraction of 0.576 × Anthropic's, 0.28 at 0.48 (0.23-0.35 over Anthropic's range), widened for survey error. General spend = 16.3 × (1 − 0.28) = US$11.7bn (low 8.7 × 0.64 = 5.57bn; high 29 × 0.8 = 23.2bn). Part total US$48.48bn. Share 24.2% (range 13.2-38.7%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: undated (the page shows none)."},"china-api":{"participates":false,"evidence":{"url":"https://developers.openai.com/api/docs/supported-countries","title":"Supported countries and territories","publisher":"OpenAI","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T2","locator":"Mainland China is not on the list; the page shows no date"}}},"screening":{"coding-agents":{"levelLow":2,"levelHigh":4},"general-reasoning":{"levelLow":2,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c636056d5a93c71b0f59276a4691dbd06387705a29a4b27c528de755c0185bbe"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c636056d5a93c71b0f59276a4691dbd06387705a29a4b27c528de755c0185bbe"}},"activityGrowth":{"current":{"low":49000000000,"central":49000000000,"high":49000000000,"unit":"USD","basis":"Annualised revenue run-rate excluding sales through cloud partners, all products (latest month × 12 or latest quarter × 4), US dollars","period":"September 2026 run-rate","periodEnd":"2026-09-30","source":{"url":"https://www.zawya.com/en/special-coverage/openais-september-annualized-revenue-nears-50bln-less-than-previously-indicated-source-says-1586243","title":"OpenAI's September annualized revenue nears $50bln, less than previously indicated, source says","publisher":"Reuters (via Zawya)","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T3","locator":"Annualised revenue 'almost $50 billion' in September; excludes sales through cloud partners such as AWS and Google Cloud, which Anthropic includes; the page shows no date"},"note":"OpenAI's annualised revenue was 'almost $50 billion' at the end of September (Reuters via Zawya; the page shows no date); taken as 49. It excludes sales through cloud partners. Total revenue is used as the measure of API activity because OpenAI gives no API figure."},"prior":{"low":20000000000,"central":20000000000,"high":22000000000,"unit":"USD","basis":"Annualised revenue run-rate excluding sales through cloud partners, all products (latest month × 12 or latest quarter × 4), US dollars","period":"End-2025 run-rate","periodEnd":"2025-12-31","source":{"url":"https://finance.yahoo.com/news/openai-cfo-says-annualized-revenue-173519097.html","title":"OpenAI CFO says annualized revenue crosses $20 billion in 2025","publisher":"Reuters (via Yahoo Finance)","publishedDate":"2026-01-19","accessedDate":"2026-10-22","tier":"T2","locator":"CFO Sarah Friar's blog post: annualised revenue above $20bn in 2025, $6bn in 2024"},"note":"OpenAI's CFO said annualised revenue passed $20bn in 2025 (blog post, quoted by Reuters on 19 Jan 2026; the blog itself returned a 403 and was not opened); taken as 20. Whether it excludes cloud-partner sales is not stated."}},"priorShares":{"coding-agents":{"low":20.1,"central":26.5,"high":33.5,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2025, annualised","periodEnd":"2025-09-30","source":{"url":"https://finance.yahoo.com/news/openai-cfo-says-annualized-revenue-173519097.html","title":"OpenAI CFO says annualized revenue crosses $20 billion in 2025","publisher":"Reuters (via Yahoo Finance)","publishedDate":"2026-01-19","accessedDate":"2026-10-22","tier":"T5","locator":"CFO Sarah Friar's blog post: annualised revenue above $20bn in 2025, $6bn in 2024"},"note":"5LayerCake estimate (T5). Prior share: OpenAI's annualised revenue was about US$14bn (12-16) at the end of September 2025, an interpolation (5LayerCake, T5) between the US$10bn it reported in June 2025 (Reuters, not re-opened) and the above US$20bn of its CFO at end-2025 (Reuters via Yahoo Finance, opened); OpenAI's spend then = today's US$16.3bn × 14 ÷ 49 = US$4.7bn (3.3-6.5), assuming the API share and the cloud-partner share were unchanged. The market total for Q3 2025 is the same frame as in Anthropic's note (6.02 times market growth for companies without evidence). Coding fractions are assumed unchanged. Spend in the part = 4.66 × 0.28 = US$1.3bn; part total US$4.93bn (central). Share 26.5% (range 20.1-33.5%). The current share (13.6%) is on the same basis, so the change is -12.9 points."},"general-reasoning":{"low":30.5,"central":38.5,"high":46.7,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2025, annualised","periodEnd":"2025-09-30","source":{"url":"https://finance.yahoo.com/news/openai-cfo-says-annualized-revenue-173519097.html","title":"OpenAI CFO says annualized revenue crosses $20 billion in 2025","publisher":"Reuters (via Yahoo Finance)","publishedDate":"2026-01-19","accessedDate":"2026-10-22","tier":"T5","locator":"CFO Sarah Friar's blog post: annualised revenue above $20bn in 2025, $6bn in 2024"},"note":"5LayerCake estimate (T5). Prior share: OpenAI's annualised revenue was about US$14bn (12-16) at the end of September 2025, an interpolation (5LayerCake, T5) between the US$10bn it reported in June 2025 (Reuters, not re-opened) and the above US$20bn of its CFO at end-2025 (Reuters via Yahoo Finance, opened); OpenAI's spend then = today's US$16.3bn × 14 ÷ 49 = US$4.7bn (3.3-6.5), assuming the API share and the cloud-partner share were unchanged. The market total for Q3 2025 is the same frame as in Anthropic's note (6.02 times market growth for companies without evidence). Coding fractions are assumed unchanged. Spend in the part = 4.66 × 0.72 = US$3.35bn; part total US$8.702bn (central). Share 38.5% (range 30.5-46.7%). The current share (24.2%) is on the same basis, so the change is -14.3 points."}}},"google-deepmind":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"google-deepmind","name":"Google DeepMind","kind":"unit","parentId":"alphabet","commercialActivityVerified":true,"shares":{"coding-agents":{"low":2.24,"central":6.44,"high":13.6,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q2 2026, annualised","periodEnd":"2026-06-30","source":{"url":"https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/","title":"Q2 2026 earnings call: Remarks from our CEO","publisher":"Google (Alphabet)","publishedDate":"2026-07-22","accessedDate":"2026-10-22","tier":"T5","locator":"'Our model APIs are now processing approximately 22 billion tokens per minute. That's up from 16 billion just a quarter ago.' Cloud backlog $514bn; no Gemini API revenue"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Google DeepMind: Alphabet's CEO said on 22 July 2026 that Google's model APIs process about 22bn tokens a minute, up from 16bn a quarter earlier (T2; Q3 2026 results are not yet out): 22bn × 525,600 minutes = 11.6 quadrillion tokens a year. Google does not report Gemini API revenue, so at an assumed blended price of US$0.69 per million tokens (0.35-1.30; Flash-heavy mix, and the count includes free-tier and internal traffic) spend is US$8bn (4-15). Unit figure, never Alphabet's totals. Coding fraction 0.27 (0.18-0.33): Menlo (T4, opened) gives Anthropic 54% and OpenAI 21% of coding spend; sharing the remaining 25% between Google and others in proportion to their 21% and 12% of all spend gives Google 15.9% of coding, a coding fraction of 0.27 against 0.356 for the market as a whole. Coding spend = 8 × 0.27 = US$2.16bn (low 4 × 0.18 = 0.72bn; high 15 × 0.33 = 4.95bn). Part total US$33.54bn. Share 6.44% (range 2.24-13.6%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-07-22 (full date)."},"general-reasoning":{"low":5.91,"central":12,"high":22.4,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q2 2026, annualised","periodEnd":"2026-06-30","source":{"url":"https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/","title":"Q2 2026 earnings call: Remarks from our CEO","publisher":"Google (Alphabet)","publishedDate":"2026-07-22","accessedDate":"2026-10-22","tier":"T5","locator":"'Our model APIs are now processing approximately 22 billion tokens per minute. That's up from 16 billion just a quarter ago.' Cloud backlog $514bn; no Gemini API revenue"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Google DeepMind: Alphabet's CEO said on 22 July 2026 that Google's model APIs process about 22bn tokens a minute, up from 16bn a quarter earlier (T2; Q3 2026 results are not yet out): 22bn × 525,600 minutes = 11.6 quadrillion tokens a year. Google does not report Gemini API revenue, so at an assumed blended price of US$0.69 per million tokens (0.35-1.30; Flash-heavy mix, and the count includes free-tier and internal traffic) spend is US$8bn (4-15). Unit figure, never Alphabet's totals. Coding fraction 0.27 (0.18-0.33): Menlo (T4, opened) gives Anthropic 54% and OpenAI 21% of coding spend; sharing the remaining 25% between Google and others in proportion to their 21% and 12% of all spend gives Google 15.9% of coding, a coding fraction of 0.27 against 0.356 for the market as a whole. General spend = 8 × (1 − 0.27) = US$5.84bn (low 4 × 0.67 = 2.68bn; high 15 × 0.82 = 12.3bn). Part total US$48.48bn. Share 12% (range 5.91-22.4%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-07-22 (full date)."},"china-api":{"participates":false,"evidence":{"url":"https://ai.google.dev/gemini-api/docs/available-regions","title":"Available regions for the Gemini API","publisher":"Google","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T2","locator":"Mainland China is not on the list; the page says 'Last updated 2026-10-22 UTC'"}}},"screening":{"coding-agents":{"levelLow":2,"levelHigh":3},"general-reasoning":{"levelLow":2,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"38551bd976bc33817161b011e90308371e3d2dd985bf75d54b20c9196f225977"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"38551bd976bc33817161b011e90308371e3d2dd985bf75d54b20c9196f225977"}},"activityGrowth":{"current":{"low":22,"central":22,"high":22,"unit":"billion tokens a minute","basis":"Tokens a minute processed through Google's first-party model APIs by customers (direct API use), as stated on the quarterly earnings call","period":"Q2 2026 (call of 22 Jul 2026)","periodEnd":"2026-06-30","source":{"url":"https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/","title":"Q2 2026 earnings call: Remarks from our CEO","publisher":"Google (Alphabet)","publishedDate":"2026-07-22","accessedDate":"2026-10-22","tier":"T2","locator":"'Our model APIs are now processing approximately 22 billion tokens per minute. That's up from 16 billion just a quarter ago.' Cloud backlog $514bn; no Gemini API revenue"},"note":"Alphabet's CEO: 'Our model APIs are now processing approximately 22 billion tokens per minute.' Q3 2026 results are not yet published. Tokens are used as the measure of activity because Google gives no API revenue; falling prices make token growth overstate revenue growth."},"prior":{"low":7,"central":7,"high":7,"unit":"billion tokens a minute","basis":"Tokens a minute processed through Google's first-party model APIs by customers (direct API use), as stated on the quarterly earnings call","period":"Q3 2025 (call of 29 Oct 2025)","periodEnd":"2025-09-30","source":{"url":"https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2025/","title":"Q3 2025 earnings call: Remarks from our CEO","publisher":"Google (Alphabet)","publishedDate":"2025-10-29","accessedDate":"2026-10-22","tier":"T2","locator":"'Our first-party models, like Gemini, now process 7 billion tokens per minute, via direct API use by our customers.'"},"note":"Alphabet's CEO: 'Our first-party models, like Gemini, now process 7 billion tokens per minute, via direct API use by our customers.' Nine months before the Q2 2026 observation, the shortest window the engine allows."}},"priorShares":{"coding-agents":{"low":7.49,"central":13.9,"high":20.6,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2025, annualised","periodEnd":"2025-09-30","source":{"url":"https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2025/","title":"Q3 2025 earnings call: Remarks from our CEO","publisher":"Google (Alphabet)","publishedDate":"2025-10-29","accessedDate":"2026-10-22","tier":"T5","locator":"'Our first-party models, like Gemini, now process 7 billion tokens per minute, via direct API use by our customers.'"},"note":"5LayerCake estimate (T5). Prior share: Google's API traffic was 7bn tokens a minute on the 29 Oct 2025 call (T2) against 22bn on 22 Jul 2026, so its spend then = today's US$8bn × 7 ÷ 22 = US$2.5bn (1.3-4.1) at an unchanged price per token (5LayerCake, T5). The market total for Q3 2025 is the same frame as in Anthropic's note (6.02 times market growth for companies without evidence). Coding fractions are assumed unchanged. Observation period: Q3 2025, annualised, nine months before the Q2 2026 current share. Spend in the part = 2.55 × 0.27 = US$0.687bn; part total US$4.93bn (central). Share 13.9% (range 7.49-20.6%). The current share (6.44%) is on the same basis, so the change is -7.5 points."},"general-reasoning":{"low":12,"central":21.4,"high":30.3,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2025, annualised","periodEnd":"2025-09-30","source":{"url":"https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2025/","title":"Q3 2025 earnings call: Remarks from our CEO","publisher":"Google (Alphabet)","publishedDate":"2025-10-29","accessedDate":"2026-10-22","tier":"T5","locator":"'Our first-party models, like Gemini, now process 7 billion tokens per minute, via direct API use by our customers.'"},"note":"5LayerCake estimate (T5). Prior share: Google's API traffic was 7bn tokens a minute on the 29 Oct 2025 call (T2) against 22bn on 22 Jul 2026, so its spend then = today's US$8bn × 7 ÷ 22 = US$2.5bn (1.3-4.1) at an unchanged price per token (5LayerCake, T5). The market total for Q3 2025 is the same frame as in Anthropic's note (6.02 times market growth for companies without evidence). Coding fractions are assumed unchanged. Observation period: Q3 2025, annualised, nine months before the Q2 2026 current share. Spend in the part = 2.55 × 0.73 = US$1.86bn; part total US$8.702bn (central). Share 21.4% (range 12-30.3%). The current share (12%) is on the same basis, so the change is -9.35 points."}}},"bytedance-seed":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"bytedance-seed","name":"ByteDance Seed","kind":"unit","parentId":"bytedance","commercialActivityVerified":true,"shares":{"coding-agents":{"low":0.00895,"central":0.0894,"high":0.593,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","title":"中国MaaS市场进入高速增长期，Token经济从概念走向规模化实践 (China's MaaS market enters a high-growth phase)","publisher":"IDC","publishedDate":"2026-05-07","accessedDate":"2026-10-22","tier":"T5","locator":"2025 public-cloud MaaS revenue ranking: Volcano Engine (over 40% share), Alibaba Cloud, Baidu AI Cloud, Zhipu, China Mobile Cloud; Tencent Cloud, SenseTime, Huawei Cloud and China Telecom Cloud named as other vendors"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for ByteDance Seed: Overseas sales of Doubao and Seed models are not separately disclosed; allowance US$0.10bn (0.02-0.4) outside mainland China. Coding fraction 0.30 (0.15-0.50). Mainland China is the main part (below). Unit only, never ByteDance's totals. Coding spend = 0.1 × 0.3 = US$30m (low 0.02 × 0.15 = 3m; high 0.4 × 0.5 = 0.2bn). Part total US$33.54bn. Share 0.0894% (range 0.00895-0.593%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-05-07 (full date)."},"general-reasoning":{"low":0.0207,"central":0.144,"high":0.697,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","title":"中国MaaS市场进入高速增长期，Token经济从概念走向规模化实践 (China's MaaS market enters a high-growth phase)","publisher":"IDC","publishedDate":"2026-05-07","accessedDate":"2026-10-22","tier":"T5","locator":"2025 public-cloud MaaS revenue ranking: Volcano Engine (over 40% share), Alibaba Cloud, Baidu AI Cloud, Zhipu, China Mobile Cloud; Tencent Cloud, SenseTime, Huawei Cloud and China Telecom Cloud named as other vendors"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for ByteDance Seed: Overseas sales of Doubao and Seed models are not separately disclosed; allowance US$0.10bn (0.02-0.4) outside mainland China. Coding fraction 0.30 (0.15-0.50). Mainland China is the main part (below). Unit only, never ByteDance's totals. General spend = 0.1 × (1 − 0.3) = US$70m (low 0.02 × 0.5 = 10m; high 0.4 × 0.85 = 0.34bn). Part total US$48.48bn. Share 0.144% (range 0.0207-0.697%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-05-07 (full date)."},"china-api":{"low":21.8,"central":43.8,"high":57.5,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://kr-asia.com/bytedance-raises-volcano-engines-maas-revenue-target-on-seedance-2-0-growth","title":"ByteDance raises Volcano Engine's MaaS revenue target on Seedance 2.0 growth","publisher":"KrASIA (reporting 36Kr)","publishedDate":"2026-06-05","accessedDate":"2026-10-22","tier":"T5","locator":"Volcano Engine MaaS revenue about RMB 1.5bn in 2025 (US$221m at RMB 6.79); 2026 target RMB 15bn (raised in April 2026, from RMB 10bn at end-2025); Seedance 2.0 monthly revenue above RMB 1bn"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for ByteDance Seed: Volcano Engine's MaaS revenue was about RMB 1.5bn in 2025 (36Kr via KrASIA, T3), in line with IDC's 2025 public-cloud MaaS market of RMB 3.07bn and its 'over 40%' revenue share for Volcano Engine (T4); its 2026 target is RMB 15bn, raised in April from RMB 10bn (36Kr via KrASIA, T3). Jefferies (24 Jun 2026, via Yahoo Finance) says Seedance, a video model that counts in Specialist models, is more than half of MaaS revenue and nearly half of that is overseas. Volcano says Doubao daily token calls reached 180 trillion in June 2026 (T2, 36Kr). Calculation: MaaS run-rate RMB 15bn (10-20, the 2026 target as a proxy) × non-video share 45% (35-50%) × closed Doubao and Seed models 85% (65-95%, since Volcano also hosts open third-party models) × mainland 92% (85-97%) = RMB 5.3bn (1.9-9.2). Company revenue RMB 5.3bn (low 1.9, high 9.2); part total RMB 12.11bn. Share 43.8% (range 21.8-57.5%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-06-05 (full date); the IDC ranking (7 May 2026, https://www.idc.com/?p=114742) is the market anchor."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"502eaca0508b5bd453f4ec2e25a271f1b7e8b71b233fa581a2d2fab03209df12"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"502eaca0508b5bd453f4ec2e25a271f1b7e8b71b233fa581a2d2fab03209df12"}},"activityGrowth":{"current":{"low":180,"central":180,"high":180,"unit":"trillion tokens a day","basis":"Doubao large-model average daily token calls, all channels, as stated by Volcano Engine","period":"June 2026","periodEnd":"2026-06-30","source":{"url":"https://eu.36kr.com/en/p/3869701615172233","title":"Volcano Engine FORCE conference: Doubao daily token calls","publisher":"36Kr","publishedDate":"2026-06-26","accessedDate":"2026-10-22","tier":"T2","locator":"Volcano Engine president Tan Dai: Doubao's daily average token calls have reached 180 trillion (FORCE Primitive Power Conference on the 23rd, taken to be 23 June 2026)"},"note":"Volcano Engine's president said Doubao's daily average token calls reached 180 trillion at the FORCE conference (36Kr, 26 Jun 2026). Includes consumer-app and internal traffic as well as API calls; falling prices make token growth overstate revenue growth."},"prior":{"low":16.4,"central":16.4,"high":16.4,"unit":"trillion tokens a day","basis":"Doubao large-model average daily token calls, all channels, as stated by Volcano Engine","period":"End of May 2025","periodEnd":"2025-05-31","source":{"url":"https://news.aibase.com/news/18819","title":"Doubao large model daily token usage reaches 16.4 trillion","publisher":"AIbase","publishedDate":"2025-06-11","accessedDate":"2026-10-22","tier":"T2","locator":"As of the end of May 2025 Doubao's daily token usage reached 16.4 trillion (137 times its release level), disclosed by Volcano Engine president Tan Dai at the FORCE conference"},"note":"Volcano Engine's president said Doubao's daily token usage reached 16.4 trillion as of the end of May 2025 (FORCE conference, reported by AIbase on 11 Jun 2025). Thirteen months before the June 2026 figure. Growth is about 11 times, above the engine's +300% cap."}},"baselineRevenueUsd":{"low":221000000,"central":221000000,"high":221000000,"unit":"USD","basis":"Public-cloud MaaS revenue of Volcano Engine, calendar year, US dollars (Seedance 2.0, a video model, launched in February 2026)","period":"CY2025","periodEnd":"2025-12-31","source":{"url":"https://kr-asia.com/bytedance-raises-volcano-engines-maas-revenue-target-on-seedance-2-0-growth","title":"ByteDance raises Volcano Engine's MaaS revenue target on Seedance 2.0 growth","publisher":"KrASIA (reporting 36Kr)","publishedDate":"2026-06-05","accessedDate":"2026-10-22","tier":"T3","locator":"Volcano Engine MaaS revenue about RMB 1.5bn in 2025 (US$221m at RMB 6.79); 2026 target RMB 15bn (raised in April 2026, from RMB 10bn at end-2025); Seedance 2.0 monthly revenue above RMB 1bn"},"note":"36Kr (via KrASIA) learned that Volcano Engine's MaaS revenue was about RMB 1.5bn in 2025; KrASIA converts it at RMB 6.79 per US dollar on 4 June 2026 (US$221m). It is the cloud unit's MaaS revenue, which also covers third-party models, not Seed models alone."}},"alibaba-qwen":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"alibaba-qwen","name":"Alibaba Qwen","kind":"unit","parentId":"alibaba","commercialActivityVerified":true,"shares":{"coding-agents":{"low":0.0119,"central":0.0939,"high":0.446,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.fool.com/earnings/call-transcripts/2026/08/27/alibaba-baba-q1-2027-earnings-call-transcript/","title":"Alibaba (BABA) Q1 2027 earnings call transcript","publisher":"Alibaba Group (call transcript published by The Motley Fool)","publishedDate":"2026-08","accessedDate":"2026-10-22","tier":"T5","locator":"CEO Eddie Wu: 'the ARR of our model and application services, including MaaS, has surpassed RMB 16 billion'; an executive: proprietary models 'still account for the majority of the revenue', third-party model revenue 'is also not small'; Qwen3.8-Max weights opened. Call on 20 Aug 2026 (the URL shows 27 Aug); the page shows no day in the text read"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Alibaba Qwen: Alibaba Cloud's Model Studio lists commercial Qwen models, most linked to its Singapore console (T2), so some sales are outside mainland China; Alibaba gives no overseas split. Allowance from the ARR below: RMB 16bn × proprietary 60% × closed Qwen 45% × overseas 15% = RMB 0.65bn, US$0.09bn (0.02-0.3). Coding fraction 0.35 (0.20-0.50). Unit only, never Alibaba's totals. Coding spend = 0.09 × 0.35 = US$31.5m (low 0.02 × 0.2 = 4m; high 0.3 × 0.5 = 0.15bn). Part total US$33.54bn. Share 0.0939% (range 0.0119-0.446%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-08 (month only)."},"general-reasoning":{"low":0.0206,"central":0.121,"high":0.493,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.fool.com/earnings/call-transcripts/2026/08/27/alibaba-baba-q1-2027-earnings-call-transcript/","title":"Alibaba (BABA) Q1 2027 earnings call transcript","publisher":"Alibaba Group (call transcript published by The Motley Fool)","publishedDate":"2026-08","accessedDate":"2026-10-22","tier":"T5","locator":"CEO Eddie Wu: 'the ARR of our model and application services, including MaaS, has surpassed RMB 16 billion'; an executive: proprietary models 'still account for the majority of the revenue', third-party model revenue 'is also not small'; Qwen3.8-Max weights opened. Call on 20 Aug 2026 (the URL shows 27 Aug); the page shows no day in the text read"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Alibaba Qwen: Alibaba Cloud's Model Studio lists commercial Qwen models, most linked to its Singapore console (T2), so some sales are outside mainland China; Alibaba gives no overseas split. Allowance from the ARR below: RMB 16bn × proprietary 60% × closed Qwen 45% × overseas 15% = RMB 0.65bn, US$0.09bn (0.02-0.3). Coding fraction 0.35 (0.20-0.50). Unit only, never Alibaba's totals. General spend = 0.09 × (1 − 0.35) = US$58.5m (low 0.02 × 0.5 = 10m; high 0.3 × 0.8 = 0.24bn). Part total US$48.48bn. Share 0.121% (range 0.0206-0.493%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-08 (month only)."},"china-api":{"low":15.1,"central":30.5,"high":54.1,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.fool.com/earnings/call-transcripts/2026/08/27/alibaba-baba-q1-2027-earnings-call-transcript/","title":"Alibaba (BABA) Q1 2027 earnings call transcript","publisher":"Alibaba Group (call transcript published by The Motley Fool)","publishedDate":"2026-08","accessedDate":"2026-10-22","tier":"T5","locator":"CEO Eddie Wu: 'the ARR of our model and application services, including MaaS, has surpassed RMB 16 billion'; an executive: proprietary models 'still account for the majority of the revenue', third-party model revenue 'is also not small'; Qwen3.8-Max weights opened. Call on 20 Aug 2026 (the URL shows 27 Aug); the page shows no day in the text read"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for Alibaba Qwen: Alibaba says (June-quarter call, 20 Aug 2026, T2) that the ARR of its model and application services, including MaaS, passed RMB 16bn, that proprietary models still account for the majority of MaaS revenue while third-party model revenue 'is also not small', and that it opened the weights of Qwen3.8-Max in August (open models count in Open-weight). The New York Times (6 Jul 2026, via AI Weekly, T3) reported that Qwen3.6-Max and Qwen3.7-Plus had moved to closed API-only access. IDC ranks Alibaba Cloud second by 2025 MaaS revenue and token calls (T4). Calculation: ARR RMB 16bn (16-22) × proprietary share 60% (50-75%) × closed Qwen models 45% of proprietary revenue (25-65%; the Max flagship is now open, other models assumed closed, and Wan-type video and image models count elsewhere) × mainland 85% (75-92%) = RMB 3.7bn (1.5-9.9). The ARR includes application services and overseas sales. Company revenue RMB 3.7bn (low 1.5, high 9.9); part total RMB 12.11bn. Share 30.5% (range 15.1-54.1%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-08 (month only); the IDC ranking (7 May 2026, https://www.idc.com/?p=114742) is the market anchor."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c82c0d1d123cc146a079d1c941491695e7cbf29b52329187c5cba02b7295906c"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c82c0d1d123cc146a079d1c941491695e7cbf29b52329187c5cba02b7295906c"}}},"xai":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"xai","name":"xAI (SpaceXAI)","kind":"unit","parentId":"spacex","commercialActivityVerified":true,"shares":{"coding-agents":{"low":0.135,"central":0.715,"high":2.92,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.sec.gov/Archives/edgar/data/1181412/000162828026052535/spcx-20260630.htm","title":"Form 10-Q for the quarter ended 30 June 2026","publisher":"Space Exploration Technologies Corp.","publishedDate":"2026-08-04","accessedDate":"2026-10-22","tier":"T5","locator":"Note 3, revenue disaggregation: AI Solutions and Infrastructure US$2,194m (Q2 2026) against US$311m (Q2 2025); the AI segment began providing cloud services (reserved compute) to customers in 2026; no Grok API figure"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for xAI (SpaceXAI): SpaceX does not report Grok API revenue separately (the AI segment also holds X advertising and subscriptions, data licensing and, from 2026, compute rental). Its Q2 2026 10-Q (T1, https://www.sec.gov/Archives/edgar/data/1181412/000162828026052535/spcx-20260630.htm, opened) shows 'AI Solutions and Infrastructure' revenue of US$2,194m against US$311m a year earlier, but the AI segment began selling cloud services (reserved compute) in 2026, so that line cannot be read as model API spending. The June 2026 prospectus (T1, https://www.sec.gov/Archives/edgar/data/1181412/000162828026042639/spaceexplorationtechnologi.htm, local copy) attributes the Q1 2026 rise in the same line (+US$191m) to Grok and X subscriptions (+US$177m) and data licensing (+US$12m), which leaves about US$2m of the rise for everything else, including the API, so API revenue grew little year on year through March 2026 (5LayerCake inference). StartupHub (9 Jul 2026, https://www.startuphub.ai/news/figure-elon-musk-spacexai-financial-breakdown-2026-07-09, T4, opened; the page names SuperGrok subscriptions and usage-based API pricing as drivers and gives no source for the figure) puts SpaceXAI's AI-specific ARR at about US$500m; Vercel's gateway gave Grok 4.7 2.00% of launch-window spend (Vercel AI Gateway Production Index, 15 Oct 2026, https://vercel.com/blog/ai-gateway-production-index-october-2026, T2). 5LayerCake estimate US$0.8bn (0.3-2.0), lowered from the screening figure of US$1.5bn after reading the filings. Coding fraction 0.30 (0.15-0.50): Grok coding models were a large part of gateway traffic. Coding spend = 0.8 × 0.3 = US$0.24bn (low 0.3 × 0.15 = 45m; high 2 × 0.5 = 1bn). Part total US$33.54bn. Share 0.715% (range 0.135-2.92%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-08-04 (full date)."},"general-reasoning":{"low":0.312,"central":1.16,"high":3.43,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.sec.gov/Archives/edgar/data/1181412/000162828026052535/spcx-20260630.htm","title":"Form 10-Q for the quarter ended 30 June 2026","publisher":"Space Exploration Technologies Corp.","publishedDate":"2026-08-04","accessedDate":"2026-10-22","tier":"T5","locator":"Note 3, revenue disaggregation: AI Solutions and Infrastructure US$2,194m (Q2 2026) against US$311m (Q2 2025); the AI segment began providing cloud services (reserved compute) to customers in 2026; no Grok API figure"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for xAI (SpaceXAI): SpaceX does not report Grok API revenue separately (the AI segment also holds X advertising and subscriptions, data licensing and, from 2026, compute rental). Its Q2 2026 10-Q (T1, https://www.sec.gov/Archives/edgar/data/1181412/000162828026052535/spcx-20260630.htm, opened) shows 'AI Solutions and Infrastructure' revenue of US$2,194m against US$311m a year earlier, but the AI segment began selling cloud services (reserved compute) in 2026, so that line cannot be read as model API spending. The June 2026 prospectus (T1, https://www.sec.gov/Archives/edgar/data/1181412/000162828026042639/spaceexplorationtechnologi.htm, local copy) attributes the Q1 2026 rise in the same line (+US$191m) to Grok and X subscriptions (+US$177m) and data licensing (+US$12m), which leaves about US$2m of the rise for everything else, including the API, so API revenue grew little year on year through March 2026 (5LayerCake inference). StartupHub (9 Jul 2026, https://www.startuphub.ai/news/figure-elon-musk-spacexai-financial-breakdown-2026-07-09, T4, opened; the page names SuperGrok subscriptions and usage-based API pricing as drivers and gives no source for the figure) puts SpaceXAI's AI-specific ARR at about US$500m; Vercel's gateway gave Grok 4.7 2.00% of launch-window spend (Vercel AI Gateway Production Index, 15 Oct 2026, https://vercel.com/blog/ai-gateway-production-index-october-2026, T2). 5LayerCake estimate US$0.8bn (0.3-2.0), lowered from the screening figure of US$1.5bn after reading the filings. Coding fraction 0.30 (0.15-0.50): Grok coding models were a large part of gateway traffic. General spend = 0.8 × (1 − 0.3) = US$0.56bn (low 0.3 × 0.5 = 0.15bn; high 2 × 0.85 = 1.7bn). Part total US$48.48bn. Share 1.16% (range 0.312-3.43%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-08-04 (full date)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.sec.gov/Archives/edgar/data/1181412/000162828026052535/spcx-20260630.htm","title":"Form 10-Q for the quarter ended 30 June 2026","publisher":"Space Exploration Technologies Corp.","publishedDate":"2026-08-04","accessedDate":"2026-10-22","tier":"T5","locator":"Note 3, revenue disaggregation: AI Solutions and Infrastructure US$2,194m (Q2 2026) against US$311m (Q2 2025); the AI segment began providing cloud services (reserved compute) to customers in 2026; no Grok API figure"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for xAI (SpaceXAI)'s closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-08-04 (full date); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":3},"general-reasoning":{"levelLow":1,"levelHigh":3},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"49159076bdf0799d001127c5c26213061d5c1ef64924d1f384dce92faefeeeb2"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"49159076bdf0799d001127c5c26213061d5c1ef64924d1f384dce92faefeeeb2"}}},"amazon-agi":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"amazon-agi","name":"Amazon AGI (Nova)","kind":"unit","parentId":"amazon","commercialActivityVerified":true,"shares":{"coding-agents":{"low":0.0149,"central":0.179,"high":1.06,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://aws.amazon.com/nova/","title":"Amazon Nova","publisher":"Amazon Web Services","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"'Adopted by tens of thousands of customers across industries'; no usage volumes or revenue; the page shows no date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Amazon AGI (Nova): AWS says Nova is 'adopted by tens of thousands of customers' (T2) but gives no usage or revenue. 5LayerCake estimate: about 30,000 customers × an assumed US$13,000 a year each (3,000-40,000) = US$0.4bn (0.1-1.2); Nova is priced low. Unit only, never AWS's totals. Coding fraction 0.15 (0.05-0.30). Re-checked in the full pass: Amazon's Q1 and Q2 2026 calls (https://www.fool.com/earnings/call-transcripts/2026/04/29/amazon-amzn-q1-2026-earnings-call-transcript/ and the Q2 call, T2) give Bedrock figures (over 125,000 customers; Q2 spend above all earlier quarters combined) but never mention Nova; the CEO said Amazon is 'pursuing our own frontier model'. No Nova usage or revenue was found, so the estimate is kept. Coding spend = 0.4 × 0.15 = US$60m (low 0.1 × 0.05 = 5m; high 1.2 × 0.3 = 0.36bn). Part total US$33.54bn. Share 0.179% (range 0.0149-1.06%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: undated (the page shows none)."},"general-reasoning":{"low":0.145,"central":0.701,"high":2.31,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://aws.amazon.com/nova/","title":"Amazon Nova","publisher":"Amazon Web Services","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"'Adopted by tens of thousands of customers across industries'; no usage volumes or revenue; the page shows no date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Amazon AGI (Nova): AWS says Nova is 'adopted by tens of thousands of customers' (T2) but gives no usage or revenue. 5LayerCake estimate: about 30,000 customers × an assumed US$13,000 a year each (3,000-40,000) = US$0.4bn (0.1-1.2); Nova is priced low. Unit only, never AWS's totals. Coding fraction 0.15 (0.05-0.30). Re-checked in the full pass: Amazon's Q1 and Q2 2026 calls (https://www.fool.com/earnings/call-transcripts/2026/04/29/amazon-amzn-q1-2026-earnings-call-transcript/ and the Q2 call, T2) give Bedrock figures (over 125,000 customers; Q2 spend above all earlier quarters combined) but never mention Nova; the CEO said Amazon is 'pursuing our own frontier model'. No Nova usage or revenue was found, so the estimate is kept. General spend = 0.4 × (1 − 0.15) = US$0.34bn (low 0.1 × 0.7 = 70m; high 1.2 × 0.95 = 1.14bn). Part total US$48.48bn. Share 0.701% (range 0.145-2.31%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: undated (the page shows none)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://aws.amazon.com/nova/","title":"Amazon Nova","publisher":"Amazon Web Services","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"'Adopted by tens of thousands of customers across industries'; no usage volumes or revenue; the page shows no date"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for Amazon AGI (Nova)'s closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: undated (the page shows none); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"ac1f9fb3dc388181854d5814ee604aaf4b5bb3dc8ee7d34d5587b285f67db0db"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"ac1f9fb3dc388181854d5814ee604aaf4b5bb3dc8ee7d34d5587b285f67db0db"}}},"mistral":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"mistral","name":"Mistral AI","kind":"private","commercialActivityVerified":true,"shares":{"coding-agents":{"low":0.0119,"central":0.149,"high":0.734,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.calcalistech.com/ctechnews/article/r1gpvm6ofg","title":"Mistral raises about €3 billion at about €21 billion valuation","publisher":"Calcalist","publishedDate":"2026-09-08","accessedDate":"2026-10-22","tier":"T5","locator":"CFO says Mistral is 'on track to reach $1 billion in annual recurring revenue by the end of this year'; more than 125 customer companies; no split between API, licences and Le Chat"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Mistral AI: Mistral's CFO says it is on track for US$1bn ARR by the end of 2026 (Calcalist, 8 Sep 2026, T2); Mistral said in February 2026 that ARR was above US$400m, about US$20m a year earlier (TechCrunch, 4 Jul 2026, opened, T3); a straight line from US$0.4bn in February to the US$1bn target gives about US$0.8bn in September. ARR now assumed US$0.7bn (0.4-1.0). Calcalist gives no split of ARR between API, licences and Le Chat, and notes its models can be downloaded and run by customers, so closed-model API spend is assumed 30% (10-55%) of ARR: 0.7 × 0.30 = US$0.21bn, rounded to $0.2bn (0.04-0.55). Coding fraction 0.25 (0.10-0.45): Codestral and Devstral Medium. Re-checked in the full pass: no published split of ARR between API, licences and Le Chat was found (Sacra describes the revenue lines but its figures are estimates and are not used), so the 30% closed-model API assumption is kept. Coding spend = 0.2 × 0.25 = US$50m (low 0.04 × 0.1 = 4m; high 0.55 × 0.45 = 0.248bn). Part total US$33.54bn. Share 0.149% (range 0.0119-0.734%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-09-08 (full date)."},"general-reasoning":{"low":0.0455,"central":0.309,"high":1.01,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.calcalistech.com/ctechnews/article/r1gpvm6ofg","title":"Mistral raises about €3 billion at about €21 billion valuation","publisher":"Calcalist","publishedDate":"2026-09-08","accessedDate":"2026-10-22","tier":"T5","locator":"CFO says Mistral is 'on track to reach $1 billion in annual recurring revenue by the end of this year'; more than 125 customer companies; no split between API, licences and Le Chat"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Mistral AI: Mistral's CFO says it is on track for US$1bn ARR by the end of 2026 (Calcalist, 8 Sep 2026, T2); Mistral said in February 2026 that ARR was above US$400m, about US$20m a year earlier (TechCrunch, 4 Jul 2026, opened, T3); a straight line from US$0.4bn in February to the US$1bn target gives about US$0.8bn in September. ARR now assumed US$0.7bn (0.4-1.0). Calcalist gives no split of ARR between API, licences and Le Chat, and notes its models can be downloaded and run by customers, so closed-model API spend is assumed 30% (10-55%) of ARR: 0.7 × 0.30 = US$0.21bn, rounded to $0.2bn (0.04-0.55). Coding fraction 0.25 (0.10-0.45): Codestral and Devstral Medium. Re-checked in the full pass: no published split of ARR between API, licences and Le Chat was found (Sacra describes the revenue lines but its figures are estimates and are not used), so the 30% closed-model API assumption is kept. General spend = 0.2 × (1 − 0.25) = US$0.15bn (low 0.04 × 0.55 = 22m; high 0.55 × 0.9 = 0.495bn). Part total US$48.48bn. Share 0.309% (range 0.0455-1.01%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-09-08 (full date)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.calcalistech.com/ctechnews/article/r1gpvm6ofg","title":"Mistral raises about €3 billion at about €21 billion valuation","publisher":"Calcalist","publishedDate":"2026-09-08","accessedDate":"2026-10-22","tier":"T5","locator":"CFO says Mistral is 'on track to reach $1 billion in annual recurring revenue by the end of this year'; more than 125 customer companies; no split between API, licences and Le Chat"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for Mistral AI's closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-09-08 (full date); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":3},"general-reasoning":{"levelLow":1,"levelHigh":3},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"8f0c417bb3573250d33673a8b6e74bff0a7b64b13a176151b87f859a6fcf22e8"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"8f0c417bb3573250d33673a8b6e74bff0a7b64b13a176151b87f859a6fcf22e8"}},"activityGrowth":{"current":{"low":400000000,"central":400000000,"high":400000000,"unit":"USD","basis":"Annual recurring revenue (ARR), all products, US dollars","period":"February 2026 ARR","periodEnd":"2026-02-28","source":{"url":"https://techcrunch.com/2026/07/04/what-is-mistral-ai-everything-to-know-about-the-openai-competitor/","title":"What is Mistral AI? Everything to know about the OpenAI competitor","publisher":"TechCrunch","publishedDate":"2026-07-04","accessedDate":"2026-10-22","tier":"T3","locator":"Mistral said in February 2026 that ARR was 'above $400 million', about $20 million a year earlier; it claimed to be on track to surpass $1 billion in ARR this year; no API split"},"note":"Mistral said in February 2026 that ARR was 'above $400 million' (TechCrunch, 4 Jul 2026; reported earlier by the FT). Total ARR is used as the measure of API activity because Mistral gives no split between API, licences and its chat app; its CFO says it is on track for US$1bn by the end of 2026 (Calcalist, 8 Sep 2026)."},"prior":{"low":20000000,"central":20000000,"high":20000000,"unit":"USD","basis":"Annual recurring revenue (ARR), all products, US dollars","period":"February 2025 ARR","periodEnd":"2025-02-28","source":{"url":"https://techcrunch.com/2026/07/04/what-is-mistral-ai-everything-to-know-about-the-openai-competitor/","title":"What is Mistral AI? Everything to know about the OpenAI competitor","publisher":"TechCrunch","publishedDate":"2026-07-04","accessedDate":"2026-10-22","tier":"T3","locator":"Mistral said in February 2026 that ARR was 'above $400 million', about $20 million a year earlier; it claimed to be on track to surpass $1 billion in ARR this year; no API split"},"note":"The same article gives ARR of about $20 million a year earlier. Growth is about 20 times, above the engine's +300% cap."}}},"baidu-ernie":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"baidu-ernie","name":"Baidu ERNIE","kind":"unit","parentId":"baidu","commercialActivityVerified":true,"shares":{"coding-agents":{"low":0,"central":0.00596,"high":0.0596,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://amp.kr-asia.com/baidus-ai-revenue-share-holds-at-50-as-wall-street-funds-buy-in","title":"Baidu's AI revenue share holds at 50% as Wall Street funds buy in","publisher":"KrASIA","publishedDate":"2026-08-19","accessedDate":"2026-10-22","tier":"T5","locator":"Q2 2026: AI Cloud Infra revenue RMB 7.3bn (+50%), GPU cloud +283%; AI-powered businesses RMB 12.5bn, 50% of general business revenue; no Qianfan MaaS or ERNIE API figure"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Baidu ERNIE: No overseas ERNIE sales figure was found; allowance US$0.01bn (0-0.05). Coding fraction 0.20 (0.05-0.40). Unit only, never Baidu's totals. Re-checked in the full pass: no absolute Qianfan or ERNIE API figure was found in a source that could be opened (the Q2 2026 call transcript pages returned 403 or were finance portals), so the estimate is kept. Coding spend = 0.01 × 0.2 = US$2m (low 0 × 0.05 = 0m; high 0.05 × 0.4 = 20m). Part total US$33.54bn. Share 0.00596% (range 0-0.0596%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-08-19 (full date)."},"general-reasoning":{"low":0,"central":0.0165,"high":0.0979,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://amp.kr-asia.com/baidus-ai-revenue-share-holds-at-50-as-wall-street-funds-buy-in","title":"Baidu's AI revenue share holds at 50% as Wall Street funds buy in","publisher":"KrASIA","publishedDate":"2026-08-19","accessedDate":"2026-10-22","tier":"T5","locator":"Q2 2026: AI Cloud Infra revenue RMB 7.3bn (+50%), GPU cloud +283%; AI-powered businesses RMB 12.5bn, 50% of general business revenue; no Qianfan MaaS or ERNIE API figure"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Baidu ERNIE: No overseas ERNIE sales figure was found; allowance US$0.01bn (0-0.05). Coding fraction 0.20 (0.05-0.40). Unit only, never Baidu's totals. Re-checked in the full pass: no absolute Qianfan or ERNIE API figure was found in a source that could be opened (the Q2 2026 call transcript pages returned 403 or were finance portals), so the estimate is kept. General spend = 0.01 × (1 − 0.2) = US$8m (low 0 × 0.6 = 0m; high 0.05 × 0.95 = 47.5m). Part total US$48.48bn. Share 0.0165% (range 0-0.0979%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-08-19 (full date)."},"china-api":{"low":2.61,"central":7.43,"high":15.1,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","title":"中国MaaS市场进入高速增长期，Token经济从概念走向规模化实践 (China's MaaS market enters a high-growth phase)","publisher":"IDC","publishedDate":"2026-05-07","accessedDate":"2026-10-22","tier":"T5","locator":"2025 public-cloud MaaS revenue ranking: Volcano Engine (over 40% share), Alibaba Cloud, Baidu AI Cloud, Zhipu, China Mobile Cloud; Tencent Cloud, SenseTime, Huawei Cloud and China Telecom Cloud named as other vendors"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for Baidu ERNIE: IDC ranks Baidu AI Cloud third by 2025 MaaS revenue and token calls (T4). Baidu's Q2 2026 results show AI Cloud Infra revenue of RMB 7.3bn, +50% (KrASIA, T3); on the call Baidu said external token-usage revenue on Qianfan rose more than ninefold and is still a small share of AI Cloud Infra revenue (the transcript was read on a finance portal and is not cited). ERNIE 5.0 is API-only. Calculation: market RMB 22bn (18.6-28) × Baidu share 9% (4-12%) × closed 45% (25-70%) = RMB 0.9bn (0.3-2.0). Company revenue RMB 0.9bn (low 0.3, high 2); part total RMB 12.11bn. Share 7.43% (range 2.61-15.1%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source published date: 2026-05-07 (full date)."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"118eb0e83cec80c6f770916d569018376ba29703bc2f0c77727b379bc025aef6"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"118eb0e83cec80c6f770916d569018376ba29703bc2f0c77727b379bc025aef6"}}},"meta-superintelligence-labs":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"meta-superintelligence-labs","name":"Meta Superintelligence Labs","kind":"unit","parentId":"meta","commercialActivityVerified":true,"shares":{"coding-agents":{"low":0.00895,"central":0.107,"high":0.741,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.trendingtopics.eu/meta-opens-muse-spark-1-1-to-developers-and-bets-on-aggressive-pricing/","title":"Meta opens Muse Spark 1.1 to developers and bets on aggressive pricing","publisher":"Trending Topics","publishedDate":"2026-07-09","accessedDate":"2026-10-22","tier":"T5","locator":"Meta Model API public preview from 9 July 2026, US developers only, $1.25 per million input and $4.25 per million output tokens; no usage or revenue disclosed; page dated 9 July 2026 (full date)"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Meta Superintelligence Labs: Muse Spark 1.1 became a paid public-preview API on 9 July 2026 for US developers at $1.25/$4.25 per million tokens (Trending Topics, T3); no usage or revenue is disclosed. Vercel's gateway gave the Muse Spark 1.3 Contributor tier 0.1% of launch-window spend; scaling by Anthropic's API spend ÷ its 43.9% gateway share (52 ÷ 0.439 = US$118bn per 100% of gateway spend) gives US$0.12bn. Range 0.02-0.5. Unit only, never Meta's totals. Coding fraction 0.30 (0.15-0.50): Meta says 1.1 gains focus on coding and multi-step agent tasks. Re-checked in the full pass (22 Oct 2026): coverage of the 9 July launch (The Verge, Bloomberg and Axios, via Qz and Trending Topics) adds only pricing, US-only public preview and $20 in free credits; no usage or revenue was found, so the estimate is kept. Coding spend = 0.12 × 0.3 = US$36m (low 0.02 × 0.15 = 3m; high 0.5 × 0.5 = 0.25bn). Part total US$33.54bn. Share 0.107% (range 0.00895-0.741%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-07-09 (full date)."},"general-reasoning":{"low":0.0207,"central":0.173,"high":0.87,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.trendingtopics.eu/meta-opens-muse-spark-1-1-to-developers-and-bets-on-aggressive-pricing/","title":"Meta opens Muse Spark 1.1 to developers and bets on aggressive pricing","publisher":"Trending Topics","publishedDate":"2026-07-09","accessedDate":"2026-10-22","tier":"T5","locator":"Meta Model API public preview from 9 July 2026, US developers only, $1.25 per million input and $4.25 per million output tokens; no usage or revenue disclosed; page dated 9 July 2026 (full date)"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Meta Superintelligence Labs: Muse Spark 1.1 became a paid public-preview API on 9 July 2026 for US developers at $1.25/$4.25 per million tokens (Trending Topics, T3); no usage or revenue is disclosed. Vercel's gateway gave the Muse Spark 1.3 Contributor tier 0.1% of launch-window spend; scaling by Anthropic's API spend ÷ its 43.9% gateway share (52 ÷ 0.439 = US$118bn per 100% of gateway spend) gives US$0.12bn. Range 0.02-0.5. Unit only, never Meta's totals. Coding fraction 0.30 (0.15-0.50): Meta says 1.1 gains focus on coding and multi-step agent tasks. Re-checked in the full pass (22 Oct 2026): coverage of the 9 July launch (The Verge, Bloomberg and Axios, via Qz and Trending Topics) adds only pricing, US-only public preview and $20 in free credits; no usage or revenue was found, so the estimate is kept. General spend = 0.12 × (1 − 0.3) = US$84m (low 0.02 × 0.5 = 10m; high 0.5 × 0.85 = 0.425bn). Part total US$48.48bn. Share 0.173% (range 0.0207-0.87%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-07-09 (full date)."},"china-api":{"participates":false,"evidence":{"url":"https://www.trendingtopics.eu/meta-opens-muse-spark-1-1-to-developers-and-bets-on-aggressive-pricing/","title":"Meta opens Muse Spark 1.1 to developers and bets on aggressive pricing","publisher":"Trending Topics","publishedDate":"2026-07-09","accessedDate":"2026-10-22","tier":"T3","locator":"Access is limited to US developers at launch ('Meta Model API', public preview from 9 July 2026); page dated 9 July 2026 (full date)"}}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":3},"general-reasoning":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"d6e8a822d05cc5e06629f20781c9ae9ed2d4d62be78491571611c3d86ff00db1"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"d6e8a822d05cc5e06629f20781c9ae9ed2d4d62be78491571611c3d86ff00db1"}}},"tencent-hy":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"tencent-hy","name":"Tencent Hy","kind":"unit","parentId":"tencent","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0.000745,"central":0.0179,"high":0.143,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.thestar.com.my/tech/tech-news/2026/08/12/tencent-q2-revenue-climbs-11-on-ai-driven-ad-gains-but-profit-falls-short","title":"Tencent Q2 revenue climbs 11% on AI-driven ad gains, but profit falls short","publisher":"Reuters (via The Star)","publishedDate":"2026-08-12","accessedDate":"2026-10-22","tier":"T5","locator":"Tencent released Hy3 in July 2026 and opened it to users worldwide; no API revenue, open-weight status or Tencent Cloud AI figure"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Tencent Hy: No overseas Hy API sales figure was found; allowance US$0.03bn (0.005-0.12). Coding fraction 0.20 (0.05-0.40). Unit only, never Tencent's totals. Re-checked in the full pass: Tencent's Q2 2026 slides say Hy3's daily token usage is about seven times the preview's and that it ranks in the top three on OpenRouter by tokens (T2); no API revenue was disclosed, and token rank on OpenRouter says little about paid spend, so the allowance is kept. Coding spend = 0.03 × 0.2 = US$6m (low 0.005 × 0.05 = 0.25m; high 0.12 × 0.4 = 48m). Part total US$33.54bn. Share 0.0179% (range 0.000745-0.143%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-08-12 (full date)."},"general-reasoning":{"low":0.00619,"central":0.0495,"high":0.235,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.thestar.com.my/tech/tech-news/2026/08/12/tencent-q2-revenue-climbs-11-on-ai-driven-ad-gains-but-profit-falls-short","title":"Tencent Q2 revenue climbs 11% on AI-driven ad gains, but profit falls short","publisher":"Reuters (via The Star)","publishedDate":"2026-08-12","accessedDate":"2026-10-22","tier":"T5","locator":"Tencent released Hy3 in July 2026 and opened it to users worldwide; no API revenue, open-weight status or Tencent Cloud AI figure"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Tencent Hy: No overseas Hy API sales figure was found; allowance US$0.03bn (0.005-0.12). Coding fraction 0.20 (0.05-0.40). Unit only, never Tencent's totals. Re-checked in the full pass: Tencent's Q2 2026 slides say Hy3's daily token usage is about seven times the preview's and that it ranks in the top three on OpenRouter by tokens (T2); no API revenue was disclosed, and token rank on OpenRouter says little about paid spend, so the allowance is kept. General spend = 0.03 × (1 − 0.2) = US$24m (low 0.005 × 0.6 = 3m; high 0.12 × 0.95 = 0.114bn). Part total US$48.48bn. Share 0.0495% (range 0.00619-0.235%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-08-12 (full date)."},"china-api":{"low":0.847,"central":3.3,"high":9.29,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","title":"中国MaaS市场进入高速增长期，Token经济从概念走向规模化实践 (China's MaaS market enters a high-growth phase)","publisher":"IDC","publishedDate":"2026-05-07","accessedDate":"2026-10-22","tier":"T5","locator":"2025 public-cloud MaaS revenue ranking: Volcano Engine (over 40% share), Alibaba Cloud, Baidu AI Cloud, Zhipu, China Mobile Cloud; Tencent Cloud, SenseTime, Huawei Cloud and China Telecom Cloud named as other vendors"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for Tencent Hy: IDC names Tencent Cloud among MaaS vendors without a share (T4). Tencent released Hy3 in July 2026 and opened it to users worldwide (Reuters via The Star, T3); no model API revenue is disclosed, and no Tencent Cloud page was readable in this pass (the one opened showed 3D and translation APIs only). Calculation: market RMB 22bn (18.6-28) × Tencent share 3% (1-6%) × closed 60% (30-100%) = RMB 0.4bn (0.1-1.2). Company revenue RMB 0.4bn (low 0.1, high 1.2); part total RMB 12.11bn. Share 3.3% (range 0.847-9.29%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source published date: 2026-05-07 (full date)."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"74044e98406880214a51b4d5102b6441b9495dc1364f6d61422611d4ee4d4522"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"74044e98406880214a51b4d5102b6441b9495dc1364f6d61422611d4ee4d4522"}}},"cohere":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"cohere","name":"Cohere","kind":"private","commercialActivityVerified":true,"shares":{"coding-agents":{"low":0,"central":0.00477,"high":0.0745,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.techcrunch.com/2026/02/13/coheres-240m-year-sets-stage-for-ipo/","title":"Cohere's $240M year sets stage for IPO","publisher":"TechCrunch","publishedDate":"2026-02-13","accessedDate":"2026-10-22","tier":"T5","locator":"'surpassed its $200 million annual recurring revenue target in 2025, hitting $240 million' (investor memo, per CNBC); quarter-over-quarter growth of more than 50% throughout the year; no API/private-deployment split; no 2026 figure"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Cohere: Cohere's investor memo, reported by CNBC and TechCrunch (13 Feb 2026, T3; Cohere declined to comment), says it surpassed its US$200m ARR target in 2025, hitting US$240m, with sales growing more than 50% quarter on quarter through 2025; the US$240m is a 2025 figure, not a February 2026 rate. The Logic (13 Feb 2026, T3) reports the same memo and says Cohere sells mainly to large companies in regulated industries. No 2026 revenue figure from Cohere was found. ARR now (Q3 2026) is a 5LayerCake judgement of US$0.4bn (0.24-0.7); the low end equals the 2025 figure (no growth since). Closed-model API spend (Embed, Rerank, Command through the API and Bedrock) is a judgement of 20% (8-35%) of ARR: US$0.08bn (0.02-0.25). Sacra (https://sacra.com/c/cohere/, T4, undated, no method given) estimates that about 85% of Cohere's revenue comes from private deployments, which leaves about 15% for other sales and sits inside the 8-35% range; TechCrunch and The Logic give no API/private-deployment split. Coding fraction 0.02 (0-0.10): enterprise search and retrieval models. Coding spend = 0.08 × 0.02 = US$1.6m (low 0.02 × 0 = 0m; high 0.25 × 0.1 = 25m). Part total US$33.54bn. Share 0.00477% (range 0-0.0745%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-02-13 (full date)."},"general-reasoning":{"low":0.0372,"central":0.162,"high":0.514,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.techcrunch.com/2026/02/13/coheres-240m-year-sets-stage-for-ipo/","title":"Cohere's $240M year sets stage for IPO","publisher":"TechCrunch","publishedDate":"2026-02-13","accessedDate":"2026-10-22","tier":"T5","locator":"'surpassed its $200 million annual recurring revenue target in 2025, hitting $240 million' (investor memo, per CNBC); quarter-over-quarter growth of more than 50% throughout the year; no API/private-deployment split; no 2026 figure"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Cohere: Cohere's investor memo, reported by CNBC and TechCrunch (13 Feb 2026, T3; Cohere declined to comment), says it surpassed its US$200m ARR target in 2025, hitting US$240m, with sales growing more than 50% quarter on quarter through 2025; the US$240m is a 2025 figure, not a February 2026 rate. The Logic (13 Feb 2026, T3) reports the same memo and says Cohere sells mainly to large companies in regulated industries. No 2026 revenue figure from Cohere was found. ARR now (Q3 2026) is a 5LayerCake judgement of US$0.4bn (0.24-0.7); the low end equals the 2025 figure (no growth since). Closed-model API spend (Embed, Rerank, Command through the API and Bedrock) is a judgement of 20% (8-35%) of ARR: US$0.08bn (0.02-0.25). Sacra (https://sacra.com/c/cohere/, T4, undated, no method given) estimates that about 85% of Cohere's revenue comes from private deployments, which leaves about 15% for other sales and sits inside the 8-35% range; TechCrunch and The Logic give no API/private-deployment split. Coding fraction 0.02 (0-0.10): enterprise search and retrieval models. General spend = 0.08 × (1 − 0.02) = US$78.4m (low 0.02 × 0.9 = 18m; high 0.25 × 1 = 0.25bn). Part total US$48.48bn. Share 0.162% (range 0.0372-0.514%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-02-13 (full date)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://thelogic.co/briefing/cohere-tells-investors-its-revenues-have-hit-us240m-rate-report/","title":"Cohere tells investors its revenues have hit US$240M rate","publisher":"The Logic","publishedDate":"2026-02-13","accessedDate":"2026-10-22","tier":"T5","locator":"2025 ARR 'topped its US$200 million target' (CNBC memo); sales grew more than 50% quarter on quarter; crossed US$100m annualised in May 2025; no API/private-deployment split"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for Cohere's closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-02-13 (full date); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":3},"general-reasoning":{"levelLow":1,"levelHigh":3},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"ed438618c3d596355dd00ef75a0cdc5ee29b49e761b556ea46c3711c01afd4c5"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"ed438618c3d596355dd00ef75a0cdc5ee29b49e761b556ea46c3711c01afd4c5"}},"activityGrowth":{"current":{"low":240000000,"central":240000000,"high":240000000,"unit":"USD","basis":"Annual recurring revenue (ARR), all products, US dollars","period":"End of 2025 (ARR hit in 2025; month not stated)","periodEnd":"2025-12-31","source":{"url":"https://www.techcrunch.com/2026/02/13/coheres-240m-year-sets-stage-for-ipo/","title":"Cohere's $240M year sets stage for IPO","publisher":"TechCrunch","publishedDate":"2026-02-13","accessedDate":"2026-10-22","tier":"T3","locator":"'surpassed its $200 million annual recurring revenue target in 2025, hitting $240 million' (investor memo, per CNBC); quarter-over-quarter growth of more than 50% throughout the year"},"note":"Company figure from Cohere's investor memo, reported by CNBC and TechCrunch on 13 Feb 2026; Cohere declined to comment, so the company has not confirmed it. The sources say only 'in 2025' (BetaKit: 'last year'); the end of the year (2025-12-31) is an assumption, and any date from 30 Sep 2025 on keeps the prior inside the 9-15 month window. The Logic's headline of the same day reads 'US$240M rate' but its text gives no date or detail for it."},"prior":{"low":62000000,"central":62000000,"high":62000000,"unit":"USD","basis":"Annual recurring revenue (ARR), all products, US dollars","period":"End of 2024","periodEnd":"2024-12-31","source":{"url":"https://sacra.com/c/cohere/","title":"Cohere revenue, valuation & funding","publisher":"Sacra","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T4","locator":"'up 287% year-over-year from $62 million at the end of 2024'; the page shows no publication date and gives no method for the end-2024 figure"},"note":"Sacra's own estimate (T4); the page shows no publication date and no method for the end-2024 figure. It ends exactly 12 months before the end-2025 figure. The two come from different producers (Sacra for 2024, the company memo for 2025), so the multiple 240 / 62 = 3.9 mixes them. The memo's 'more than 50% quarter on quarter throughout 2025' would compound to at least 1.5^4 = 5.06, which implies end-2024 ARR of at most about US$47m (240 / 5.06), inside the T4 band around US$62m; Sacra's higher figure makes the growth multiple the more conservative one."}}},"iflytek":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"iflytek","name":"iFlytek (Spark)","kind":"public","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00149,"high":0.0268,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://eu.36kr.com/en/p/3972575395951111","title":"iFlytek H1 2026 results","publisher":"36Kr","publishedDate":"2026-09-07","accessedDate":"2026-10-22","tier":"T5","locator":"H1 2026 AI open platform revenue RMB 3.705bn (31.87% of RMB 11.623bn, +36.01%); large-model API and MaaS platform revenue about +70% year on year; no Spark API amount; page dated 7 September 2026 (full date)"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for iFlytek (Spark): No overseas Spark API sales found; allowance US$0.005bn (0-0.03). Coding fraction 0.10 (0-0.30). Spark only, never iFlytek's totals. Re-checked in the full pass: no Spark API amount was found beyond the screening evidence, so the allowance is kept. Coding spend = 0.005 × 0.1 = US$0.5m (low 0 × 0 = 0m; high 0.03 × 0.3 = 9m). Part total US$33.54bn. Share 0.00149% (range 0-0.0268%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-09-07 (full date)."},"general-reasoning":{"low":0,"central":0.00928,"high":0.0618,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://eu.36kr.com/en/p/3972575395951111","title":"iFlytek H1 2026 results","publisher":"36Kr","publishedDate":"2026-09-07","accessedDate":"2026-10-22","tier":"T5","locator":"H1 2026 AI open platform revenue RMB 3.705bn (31.87% of RMB 11.623bn, +36.01%); large-model API and MaaS platform revenue about +70% year on year; no Spark API amount; page dated 7 September 2026 (full date)"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for iFlytek (Spark): No overseas Spark API sales found; allowance US$0.005bn (0-0.03). Coding fraction 0.10 (0-0.30). Spark only, never iFlytek's totals. Re-checked in the full pass: no Spark API amount was found beyond the screening evidence, so the allowance is kept. General spend = 0.005 × (1 − 0.1) = US$4.5m (low 0 × 0.7 = 0m; high 0.03 × 1 = 30m). Part total US$48.48bn. Share 0.00928% (range 0-0.0618%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-09-07 (full date)."},"china-api":{"low":0.425,"central":3.3,"high":20.4,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://eu.36kr.com/en/p/3972575395951111","title":"iFlytek H1 2026 results","publisher":"36Kr","publishedDate":"2026-09-07","accessedDate":"2026-10-22","tier":"T5","locator":"H1 2026 AI open platform revenue RMB 3.705bn (31.87% of RMB 11.623bn, +36.01%); large-model API and MaaS platform revenue about +70% year on year; no Spark API amount; page dated 7 September 2026 (full date)"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for iFlytek (Spark): iFlytek's H1 2026 AI open platform revenue was RMB 3.705bn (31.87% of group revenue, +36%) and its large-model API and MaaS platform revenue rose about 70% (36Kr, T3); the Spark API amount is not given and the platform figure includes much else. Assumed Spark closed-model API RMB 0.4bn (0.05-3.0). Not marked verified because paying API customers are not shown. Company revenue RMB 0.4bn (low 0.05, high 3); part total RMB 12.11bn. Share 3.3% (range 0.425-20.4%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-09-07 (full date); the IDC ranking is the market anchor."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"d4ed1a4ccb9faacfdfc29d8412d0f5f48345e4e467b0e298a27bb6f84126bb2b"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"d4ed1a4ccb9faacfdfc29d8412d0f5f48345e4e467b0e298a27bb6f84126bb2b"}}},"typesafe-ai":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"typesafe-ai","name":"TypeSafe AI","kind":"private","commercialActivityVerified":true,"shares":{"coding-agents":{"low":0,"central":0.00745,"high":0.208,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","title":"AI Gateway Production Index, October 2026 (September data)","publisher":"Vercel","publishedDate":"2026-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"Anthropic 43.9% of September gateway spend; launch-window spend shares incl. Grok 4.7 2.00%, Muse Spark 1.3 Contributor 0.1%, Jev 0.15-0.3%; open-weight about 23% of spend. page dated 15 October 2026 (full date)"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for TypeSafe AI: Vercel's index calls Jev the fastest-adopted model in gateway history and gives it 0.15-0.3% of launch-window (first 14 days) spend in three places; × US$118bn per 100% of gateway spend = US$0.18-0.35bn if sustained. Launch spikes usually fade (5LayerCake judgement), so central US$0.05bn (0.005-0.35). Coding fraction 0.05 (0-0.20). Paying customers beyond the gateway are not shown. Re-checked in the full pass: the New Stack article carried only a headline in the text read, so no price, customer or revenue figure was added; the estimate is kept. Coding spend = 0.05 × 0.05 = US$2.5m (low 0.005 × 0 = 0m; high 0.35 × 0.2 = 70m). Part total US$33.54bn. Share 0.00745% (range 0-0.208%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-10-15 (full date)."},"general-reasoning":{"low":0.00826,"central":0.098,"high":0.717,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","title":"AI Gateway Production Index, October 2026 (September data)","publisher":"Vercel","publishedDate":"2026-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"Anthropic 43.9% of September gateway spend; launch-window spend shares incl. Grok 4.7 2.00%, Muse Spark 1.3 Contributor 0.1%, Jev 0.15-0.3%; open-weight about 23% of spend. page dated 15 October 2026 (full date)"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for TypeSafe AI: Vercel's index calls Jev the fastest-adopted model in gateway history and gives it 0.15-0.3% of launch-window (first 14 days) spend in three places; × US$118bn per 100% of gateway spend = US$0.18-0.35bn if sustained. Launch spikes usually fade (5LayerCake judgement), so central US$0.05bn (0.005-0.35). Coding fraction 0.05 (0-0.20). Paying customers beyond the gateway are not shown. Re-checked in the full pass: the New Stack article carried only a headline in the text read, so no price, customer or revenue figure was added; the estimate is kept. General spend = 0.05 × (1 − 0.05) = US$47.5m (low 0.005 × 0.8 = 4m; high 0.35 × 1 = 0.35bn). Part total US$48.48bn. Share 0.098% (range 0.00826-0.717%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-10-15 (full date)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","title":"AI Gateway Production Index, October 2026 (September data)","publisher":"Vercel","publishedDate":"2026-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"Anthropic 43.9% of September gateway spend; launch-window spend shares incl. Grok 4.7 2.00%, Muse Spark 1.3 Contributor 0.1%, Jev 0.15-0.3%; open-weight about 23% of spend. page dated 15 October 2026 (full date)"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for TypeSafe AI's closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-10-15 (full date); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"b63cb76c4b523dd00cfd23dd55b7cb8fbeaee6f2cacce07529df724c1439e816"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"b63cb76c4b523dd00cfd23dd55b7cb8fbeaee6f2cacce07529df724c1439e816"}}},"sber":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"sber","name":"Sber (GigaChat)","kind":"public","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0.000596,"central":0.0149,"high":0.186,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://developers.sber.ru/docs/en/gigachat/api/overview","title":"GigaChat API overview","publisher":"Sber","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"GigaChat API offered on request ('Request access to the service'); no prices, customers or usage; the page shows no date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Sber (GigaChat): Sber offers the GigaChat API on request through its developer portal (T2); prices, customers and usage are not disclosed. Allowance US$0.05bn (0.01-0.25), counted outside mainland China. Coding fraction 0.10 (0.02-0.25). Re-checked in the full pass: no GigaChat API revenue was found (Sber reports about 30bn roubles of internal efficiency gains in H1 2025, which is not sales); the allowance is kept. Coding spend = 0.05 × 0.1 = US$5m (low 0.01 × 0.02 = 0.2m; high 0.25 × 0.25 = 62.5m). Part total US$33.54bn. Share 0.0149% (range 0.000596-0.186%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: undated (the page shows none)."},"general-reasoning":{"low":0.0155,"central":0.0928,"high":0.503,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://developers.sber.ru/docs/en/gigachat/api/overview","title":"GigaChat API overview","publisher":"Sber","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"GigaChat API offered on request ('Request access to the service'); no prices, customers or usage; the page shows no date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Sber (GigaChat): Sber offers the GigaChat API on request through its developer portal (T2); prices, customers and usage are not disclosed. Allowance US$0.05bn (0.01-0.25), counted outside mainland China. Coding fraction 0.10 (0.02-0.25). Re-checked in the full pass: no GigaChat API revenue was found (Sber reports about 30bn roubles of internal efficiency gains in H1 2025, which is not sales); the allowance is kept. General spend = 0.05 × (1 − 0.1) = US$45m (low 0.01 × 0.75 = 7.5m; high 0.25 × 0.98 = 0.245bn). Part total US$48.48bn. Share 0.0928% (range 0.0155-0.503%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: undated (the page shows none)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://developers.sber.ru/docs/en/gigachat/api/overview","title":"GigaChat API overview","publisher":"Sber","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"GigaChat API offered on request ('Request access to the service'); no prices, customers or usage; the page shows no date"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for Sber (GigaChat)'s closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: undated (the page shows none); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"db60e89d0e17e3997d9faad412cf6cd765e03eaa9e834b6e026e2a1fc646101f"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"db60e89d0e17e3997d9faad412cf6cd765e03eaa9e834b6e026e2a1fc646101f"}}},"perplexity":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"perplexity","name":"Perplexity (Sonar API)","kind":"private","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00238,"high":0.0358,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://techcrunch.com/2025/01/21/perplexity-launches-sonar-an-api-for-ai-search","title":"Perplexity launches Sonar, an API for AI search","publisher":"TechCrunch","publishedDate":"2025-01-21","accessedDate":"2026-10-22","tier":"T5","locator":"Sonar and Sonar Pro API launched 21 Jan 2025; Zoom named as a customer; no revenue"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Perplexity (Sonar API): Sonar and Sonar Pro API launched 21 Jan 2025 (TechCrunch, T3) (the register notes closed fine-tunes of open base models); Zoom was a named customer. No API revenue is disclosed (consumer and seat revenue count in Applications); allowance US$0.04bn (0.01-0.12). Coding fraction 0.02 (0-0.10). Re-checked in the full pass: no Sonar API revenue was found, so the allowance is kept. Coding spend = 0.04 × 0.02 = US$0.8m (low 0.01 × 0 = 0m; high 0.12 × 0.1 = 12m). Part total US$33.54bn. Share 0.00238% (range 0-0.0358%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2025-01-21 (full date)."},"general-reasoning":{"low":0.0186,"central":0.0809,"high":0.247,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://techcrunch.com/2025/01/21/perplexity-launches-sonar-an-api-for-ai-search","title":"Perplexity launches Sonar, an API for AI search","publisher":"TechCrunch","publishedDate":"2025-01-21","accessedDate":"2026-10-22","tier":"T5","locator":"Sonar and Sonar Pro API launched 21 Jan 2025; Zoom named as a customer; no revenue"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Perplexity (Sonar API): Sonar and Sonar Pro API launched 21 Jan 2025 (TechCrunch, T3) (the register notes closed fine-tunes of open base models); Zoom was a named customer. No API revenue is disclosed (consumer and seat revenue count in Applications); allowance US$0.04bn (0.01-0.12). Coding fraction 0.02 (0-0.10). Re-checked in the full pass: no Sonar API revenue was found, so the allowance is kept. General spend = 0.04 × (1 − 0.02) = US$39.2m (low 0.01 × 0.9 = 9m; high 0.12 × 1 = 0.12bn). Part total US$48.48bn. Share 0.0809% (range 0.0186-0.247%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2025-01-21 (full date)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://techcrunch.com/2025/01/21/perplexity-launches-sonar-an-api-for-ai-search","title":"Perplexity launches Sonar, an API for AI search","publisher":"TechCrunch","publishedDate":"2025-01-21","accessedDate":"2026-10-22","tier":"T5","locator":"Sonar and Sonar Pro API launched 21 Jan 2025; Zoom named as a customer; no revenue"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for Perplexity (Sonar API)'s closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2025-01-21 (full date); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"11a734ad7b843027d04c1ba25154c8c0e5f9203d4a6823c508ea7294f4f9e9cf"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"11a734ad7b843027d04c1ba25154c8c0e5f9203d4a6823c508ea7294f4f9e9cf"}}},"naver-hyperclova":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"naver-hyperclova","name":"NAVER HyperCLOVA X","kind":"unit","parentId":"naver","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00596,"high":0.0894,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://guide.ncloud-docs.com/docs/en/clovastudio-overview","title":"CLOVA Studio overview","publisher":"NAVER Cloud","publishedDate":"2026-10-08","accessedDate":"2026-10-22","tier":"T5","locator":"CLOVA Studio is a NAVER Cloud Platform service using the HyperCLOVA X language model; no prices or revenue on the page"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for NAVER HyperCLOVA X: CLOVA Studio, NAVER Cloud's service built on HyperCLOVA X, is documented as live (page updated 8 Oct 2026, T2); prices and revenue are not disclosed. Allowance US$0.04bn (0.01-0.15). Coding fraction 0.05 (0-0.20). Unit only, never NAVER's totals. Re-checked in the full pass: no CLOVA Studio price or revenue figure was found, so the allowance is kept. Coding spend = 0.04 × 0.05 = US$2m (low 0.01 × 0 = 0m; high 0.15 × 0.2 = 30m). Part total US$33.54bn. Share 0.00596% (range 0-0.0894%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-10-08 (full date)."},"general-reasoning":{"low":0.0165,"central":0.0784,"high":0.309,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://guide.ncloud-docs.com/docs/en/clovastudio-overview","title":"CLOVA Studio overview","publisher":"NAVER Cloud","publishedDate":"2026-10-08","accessedDate":"2026-10-22","tier":"T5","locator":"CLOVA Studio is a NAVER Cloud Platform service using the HyperCLOVA X language model; no prices or revenue on the page"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for NAVER HyperCLOVA X: CLOVA Studio, NAVER Cloud's service built on HyperCLOVA X, is documented as live (page updated 8 Oct 2026, T2); prices and revenue are not disclosed. Allowance US$0.04bn (0.01-0.15). Coding fraction 0.05 (0-0.20). Unit only, never NAVER's totals. Re-checked in the full pass: no CLOVA Studio price or revenue figure was found, so the allowance is kept. General spend = 0.04 × (1 − 0.05) = US$38m (low 0.01 × 0.8 = 8m; high 0.15 × 1 = 0.15bn). Part total US$48.48bn. Share 0.0784% (range 0.0165-0.309%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-10-08 (full date)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://guide.ncloud-docs.com/docs/en/clovastudio-overview","title":"CLOVA Studio overview","publisher":"NAVER Cloud","publishedDate":"2026-10-08","accessedDate":"2026-10-22","tier":"T5","locator":"CLOVA Studio is a NAVER Cloud Platform service using the HyperCLOVA X language model; no prices or revenue on the page"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for NAVER HyperCLOVA X's closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-10-08 (full date); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"eebb4d9a90fa2ac82184437541dfaa852e2c0ef03bdee4a8a22f2081277014d8"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"eebb4d9a90fa2ac82184437541dfaa852e2c0ef03bdee4a8a22f2081277014d8"}}},"moonshot":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"moonshot","name":"Moonshot AI","kind":"private","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.0238,"high":0.268,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://aiweekly.co/alerts/moonshot-ai-tells-investors-arr-topped-1b-in-august-targets-2b-by-end-of-2026","title":"Moonshot AI Tells Investors ARR Topped $1B in August, Targets $2B by End of 2026","publisher":"AI Weekly (reporting Bloomberg)","publishedDate":"2026-09-11","accessedDate":"2026-10-22","tier":"T5","locator":"Annualised revenue passed $1bn in August 2026 (from $300m in June); Kimi K3 described as the largest open-weight model to date; no API/consumer split"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Moonshot AI: Kimi K3 is open-weight (Bloomberg via AI Weekly, 11 Sep 2026, T3), and Moonshot's reported ARR above US$1bn in August 2026 is open-weight API plus consumer (Open-weight and Consumer). Closed-model API allowance US$0.02bn (0-0.15). Coding fraction 0.40 (0.20-0.60). Re-checked in the full pass: Kimi models are open-weight; no closed variant was found, so the allowance is kept. Coding spend = 0.02 × 0.4 = US$8m (low 0 × 0.2 = 0m; high 0.15 × 0.6 = 90m). Part total US$33.54bn. Share 0.0238% (range 0-0.268%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-09-11 (full date)."},"general-reasoning":{"low":0,"central":0.0248,"high":0.247,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://aiweekly.co/alerts/moonshot-ai-tells-investors-arr-topped-1b-in-august-targets-2b-by-end-of-2026","title":"Moonshot AI Tells Investors ARR Topped $1B in August, Targets $2B by End of 2026","publisher":"AI Weekly (reporting Bloomberg)","publishedDate":"2026-09-11","accessedDate":"2026-10-22","tier":"T5","locator":"Annualised revenue passed $1bn in August 2026 (from $300m in June); Kimi K3 described as the largest open-weight model to date; no API/consumer split"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Moonshot AI: Kimi K3 is open-weight (Bloomberg via AI Weekly, 11 Sep 2026, T3), and Moonshot's reported ARR above US$1bn in August 2026 is open-weight API plus consumer (Open-weight and Consumer). Closed-model API allowance US$0.02bn (0-0.15). Coding fraction 0.40 (0.20-0.60). Re-checked in the full pass: Kimi models are open-weight; no closed variant was found, so the allowance is kept. General spend = 0.02 × (1 − 0.4) = US$12m (low 0 × 0.4 = 0m; high 0.15 × 0.8 = 0.12bn). Part total US$48.48bn. Share 0.0248% (range 0-0.247%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-09-11 (full date)."},"china-api":{"low":0,"central":1.24,"high":7.71,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://kr-asia.com/moonshot-ai-sees-overseas-revenue-surge-as-kimi-k2-5-gains-traction-abroad","title":"Moonshot AI sees overseas revenue surge as Kimi K2.5 gains traction abroad","publisher":"KrASIA","publishedDate":"2026-02-02","accessedDate":"2026-10-22","tier":"T5","locator":"'In recent discussions with investors, Moonshot said its overseas revenue has surpassed its domestic revenue'; global paying users up fourfold after the Kimi K2.5 launch; no revenue figures"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for Moonshot AI: in discussions with investors around the Kimi K2.5 launch Moonshot said its overseas revenue had passed its domestic revenue (KrASIA, 2 Feb 2026, the source above, T3; a company statement as reported by KrASIA, with no revenue figures; it is from February, not from the August report). Annualised revenue passed US$1bn in August 2026, up from US$300m in June (Bloomberg via AI Weekly, 11 Sep 2026, https://aiweekly.co/alerts/moonshot-ai-tells-investors-arr-topped-1b-in-august-targets-2b-by-end-of-2026, T3), with no split between API and consumer revenue. Kimi K3 is described as open-weight there, so only any closed variants count here. Assumed mainland closed API RMB 0.15bn (0-1.0) is a 5LayerCake judgement (T5): domestic revenue is under half of the total (KrASIA) and only closed variants of its API count. No closed Moonshot model was confirmed, so not marked verified. Company revenue RMB 0.15bn (low 0, high 1); part total RMB 12.11bn. Share 1.24% (range 0-7.71%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-02-02 (full date); the IDC ranking (7 May 2026) is the market anchor."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"6f483173006f148b762748c26ef0a8d1591f568b0a239d230dd99d45b3b5f3c4"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"6f483173006f148b762748c26ef0a8d1591f568b0a239d230dd99d45b3b5f3c4"}}},"sensetime":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"sensetime","name":"SenseTime (SenseNova)","kind":"public","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00149,"high":0.0268,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://technode.global/2026/08/27/sensetime-h1-2026-first-profit-generative-ai-revenue/","title":"SenseTime H1 2026: first profit, generative AI revenue","publisher":"TechNode","publishedDate":"2026-08-27","accessedDate":"2026-10-22","tier":"T5","locator":"H1 2026 generative AI revenue RMB 2.33bn (+28.2%, 79.9% of group); SenseCore handled over 2.4 trillion tokens a day in July, serving four external foundation-model providers; no SenseNova API revenue"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for SenseTime (SenseNova): No overseas SenseNova API sales found; allowance US$0.005bn (0-0.03). Coding fraction 0.10 (0-0.30). SenseNova only, never SenseTime's totals. Re-checked in the full pass: no SenseNova API revenue was found beyond the screening evidence, so the allowance is kept. Coding spend = 0.005 × 0.1 = US$0.5m (low 0 × 0 = 0m; high 0.03 × 0.3 = 9m). Part total US$33.54bn. Share 0.00149% (range 0-0.0268%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-08-27 (full date)."},"general-reasoning":{"low":0,"central":0.00928,"high":0.0618,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://technode.global/2026/08/27/sensetime-h1-2026-first-profit-generative-ai-revenue/","title":"SenseTime H1 2026: first profit, generative AI revenue","publisher":"TechNode","publishedDate":"2026-08-27","accessedDate":"2026-10-22","tier":"T5","locator":"H1 2026 generative AI revenue RMB 2.33bn (+28.2%, 79.9% of group); SenseCore handled over 2.4 trillion tokens a day in July, serving four external foundation-model providers; no SenseNova API revenue"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for SenseTime (SenseNova): No overseas SenseNova API sales found; allowance US$0.005bn (0-0.03). Coding fraction 0.10 (0-0.30). SenseNova only, never SenseTime's totals. Re-checked in the full pass: no SenseNova API revenue was found beyond the screening evidence, so the allowance is kept. General spend = 0.005 × (1 − 0.1) = US$4.5m (low 0 × 0.7 = 0m; high 0.03 × 1 = 30m). Part total US$48.48bn. Share 0.00928% (range 0-0.0618%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-08-27 (full date)."},"china-api":{"low":0.42,"central":2.06,"high":6.32,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","title":"中国MaaS市场进入高速增长期，Token经济从概念走向规模化实践 (China's MaaS market enters a high-growth phase)","publisher":"IDC","publishedDate":"2026-05-07","accessedDate":"2026-10-22","tier":"T5","locator":"2025 public-cloud MaaS revenue ranking: Volcano Engine (over 40% share), Alibaba Cloud, Baidu AI Cloud, Zhipu, China Mobile Cloud; Tencent Cloud, SenseTime, Huawei Cloud and China Telecom Cloud named as other vendors"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for SenseTime (SenseNova): IDC names SenseTime for token calls and private deployment (T4). SenseTime's H1 2026 generative AI revenue was RMB 2.33bn (+28.2%), mostly deployments and compute; SenseCore handled over 2.4 trillion tokens a day in July, partly for four external model providers (TechNode, T2). No SenseNova API revenue is disclosed; assumed RMB 0.25bn (0.05-0.8) for closed SenseNova models sold by API. Not marked verified. Company revenue RMB 0.25bn (low 0.05, high 0.8); part total RMB 12.11bn. Share 2.06% (range 0.42-6.32%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source published date: 2026-05-07 (full date)."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"5318690e3a782d6c31253a1b2b01ecbf296cb14e7fd422e13870c8b466bca864"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"5318690e3a782d6c31253a1b2b01ecbf296cb14e7fd422e13870c8b466bca864"}}},"microsoft-ai":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"microsoft-ai","name":"Microsoft AI (MAI)","kind":"unit","parentId":"microsoft","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0.00298,"central":0.0358,"high":0.357,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://pureai.com/articles/2026/06/05/microsoft-unveils-mai-thinking1-as-it-expands-its-inhouse-ai-model-strategy.aspx","title":"Microsoft unveils MAI-Thinking-1 as it expands its in-house AI model strategy","publisher":"Pure AI","publishedDate":"2026-06-05","accessedDate":"2026-10-22","tier":"T5","locator":"MAI-Thinking-1 in private preview through Microsoft Foundry; OpenRouter, Fireworks and Baseten listed as future availability; no pricing, usage or revenue"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Microsoft AI (MAI): MAI-Thinking-1 is in private preview through Microsoft Foundry (Pure AI, 5 Jun 2026, T3); pricing and usage are not disclosed, so no paying customer is shown. Allowance US$0.03bn (0.005-0.2). OpenAI models sold on Azure count for OpenAI. Coding fraction 0.40 (0.20-0.60): Microsoft says the model is competitive on coding benchmarks. Re-checked in the full pass: Azure's catalog lists MAI-Thinking-1 at US$2 and US$8 per million input and output tokens (a pricing listing, not Microsoft's own statement), but no usage, customer or revenue figure was found and general availability was not confirmed, so the allowance is kept. Coding spend = 0.03 × 0.4 = US$12m (low 0.005 × 0.2 = 1m; high 0.2 × 0.6 = 0.12bn). Part total US$33.54bn. Share 0.0358% (range 0.00298-0.357%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-06-05 (full date)."},"general-reasoning":{"low":0.00413,"central":0.0371,"high":0.329,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://pureai.com/articles/2026/06/05/microsoft-unveils-mai-thinking1-as-it-expands-its-inhouse-ai-model-strategy.aspx","title":"Microsoft unveils MAI-Thinking-1 as it expands its in-house AI model strategy","publisher":"Pure AI","publishedDate":"2026-06-05","accessedDate":"2026-10-22","tier":"T5","locator":"MAI-Thinking-1 in private preview through Microsoft Foundry; OpenRouter, Fireworks and Baseten listed as future availability; no pricing, usage or revenue"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Microsoft AI (MAI): MAI-Thinking-1 is in private preview through Microsoft Foundry (Pure AI, 5 Jun 2026, T3); pricing and usage are not disclosed, so no paying customer is shown. Allowance US$0.03bn (0.005-0.2). OpenAI models sold on Azure count for OpenAI. Coding fraction 0.40 (0.20-0.60): Microsoft says the model is competitive on coding benchmarks. Re-checked in the full pass: Azure's catalog lists MAI-Thinking-1 at US$2 and US$8 per million input and output tokens (a pricing listing, not Microsoft's own statement), but no usage, customer or revenue figure was found and general availability was not confirmed, so the allowance is kept. General spend = 0.03 × (1 − 0.4) = US$18m (low 0.005 × 0.4 = 2m; high 0.2 × 0.8 = 0.16bn). Part total US$48.48bn. Share 0.0371% (range 0.00413-0.329%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-06-05 (full date)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://pureai.com/articles/2026/06/05/microsoft-unveils-mai-thinking1-as-it-expands-its-inhouse-ai-model-strategy.aspx","title":"Microsoft unveils MAI-Thinking-1 as it expands its in-house AI model strategy","publisher":"Pure AI","publishedDate":"2026-06-05","accessedDate":"2026-10-22","tier":"T5","locator":"MAI-Thinking-1 in private preview through Microsoft Foundry; OpenRouter, Fireworks and Baseten listed as future availability; no pricing, usage or revenue"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for Microsoft AI (MAI)'s closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-06-05 (full date); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"1f4c8880c8b1ffa8c37d1353df71dc478ac7a70320342aa1be0e14f6bcf96397"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"1f4c8880c8b1ffa8c37d1353df71dc478ac7a70320342aa1be0e14f6bcf96397"}}},"china-mobile":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"china-mobile","name":"China Mobile (Jiutian)","kind":"public","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.000298,"high":0.00894,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://telecomlead.com/5g/china-mobile-revenue-falls-1-1-to-rmb-538-bn-as-ai-cloud-and-5g-investment-accelerates-127284","title":"China Mobile revenue falls 1.1% to RMB 538 bn as AI cloud and 5G investment accelerates","publisher":"TelecomLead","publishedDate":"2026-08-13","accessedDate":"2026-10-22","tier":"T5","locator":"H1 2026: AI services revenue RMB 49.34bn (+0.9%); Jiutian large-model revenue +12.8%, no absolute figure; no token or MaaS figure; page dated 13 August 2026 (full date)"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for China Mobile (Jiutian): No overseas Jiutian sales found; allowance US$0.001bn (0-0.01). Coding fraction 0.10 (0-0.30). Jiutian only, never China Mobile's totals; its serving of other developers' models counts in Inference serving. Coding spend = 0.001 × 0.1 = US$0.1m (low 0 × 0 = 0m; high 0.01 × 0.3 = 3m). Part total US$33.54bn. Share 0.000298% (range 0-0.00894%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-08-13 (full date)."},"general-reasoning":{"low":0,"central":0.00186,"high":0.0206,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://telecomlead.com/5g/china-mobile-revenue-falls-1-1-to-rmb-538-bn-as-ai-cloud-and-5g-investment-accelerates-127284","title":"China Mobile revenue falls 1.1% to RMB 538 bn as AI cloud and 5G investment accelerates","publisher":"TelecomLead","publishedDate":"2026-08-13","accessedDate":"2026-10-22","tier":"T5","locator":"H1 2026: AI services revenue RMB 49.34bn (+0.9%); Jiutian large-model revenue +12.8%, no absolute figure; no token or MaaS figure; page dated 13 August 2026 (full date)"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for China Mobile (Jiutian): No overseas Jiutian sales found; allowance US$0.001bn (0-0.01). Coding fraction 0.10 (0-0.30). Jiutian only, never China Mobile's totals; its serving of other developers' models counts in Inference serving. General spend = 0.001 × (1 − 0.1) = US$0.9m (low 0 × 0.7 = 0m; high 0.01 × 1 = 10m). Part total US$48.48bn. Share 0.00186% (range 0-0.0206%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-08-13 (full date)."},"china-api":{"low":0.418,"central":1.65,"high":4.8,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","title":"中国MaaS市场进入高速增长期，Token经济从概念走向规模化实践 (China's MaaS market enters a high-growth phase)","publisher":"IDC","publishedDate":"2026-05-07","accessedDate":"2026-10-22","tier":"T5","locator":"2025 public-cloud MaaS revenue ranking: Volcano Engine (over 40% share), Alibaba Cloud, Baidu AI Cloud, Zhipu, China Mobile Cloud; Tencent Cloud, SenseTime, Huawei Cloud and China Telecom Cloud named as other vendors"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for China Mobile (Jiutian): IDC ranks China Mobile Cloud fifth by 2025 MaaS revenue and token calls (T4). China Mobile's H1 2026 Jiutian large-model revenue rose 12.8% with no absolute figure (TelecomLead, T3). Most of its MaaS revenue is other developers' models (Inference serving); own closed Jiutian models are assumed 30% (10-60%) of a 3% (1-5%) share of the RMB 22bn market: RMB 0.2bn (0.05-0.6). Not marked verified. Company revenue RMB 0.2bn (low 0.05, high 0.6); part total RMB 12.11bn. Share 1.65% (range 0.418-4.8%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source published date: 2026-05-07 (full date)."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"e50e510f82936b53ab74c6c00fc78aa7e95c5524769753b8fe4656d99366401c"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"e50e510f82936b53ab74c6c00fc78aa7e95c5524769753b8fe4656d99366401c"}}},"zhipu":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"zhipu","name":"Zhipu AI (Z.ai)","kind":"public","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00596,"high":0.0894,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://whbl.com/2026/03/31/chinas-zhipu-posts-132-rise-in-annual-revenue-on-ai-boom/","title":"China's Zhipu posts 132% rise in annual revenue on AI boom","publisher":"Reuters (via WHBL)","publishedDate":"2026-03-31","accessedDate":"2026-10-22","tier":"T5","locator":"2025 revenue: on-premise RMB 533.9m; cloud-based revenue from API services RMB 190.4m (+292%); the page does not say whether the GLM models are open or closed"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Zhipu AI (Z.ai): Zhipu's flagship GLM models are open-weight (card; Open-weight sublayer); no closed overseas model sales found. Allowance US$0.005bn (0-0.05). Coding fraction 0.40 (0.20-0.60). Re-checked in the full pass: the GLM flagships are open-weight, so their API revenue belongs to Open-weight; no closed Zhipu model was found, so the allowance is kept. Coding spend = 0.005 × 0.4 = US$2m (low 0 × 0.2 = 0m; high 0.05 × 0.6 = 30m). Part total US$33.54bn. Share 0.00596% (range 0-0.0894%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-03-31 (full date)."},"general-reasoning":{"low":0,"central":0.00619,"high":0.0824,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://whbl.com/2026/03/31/chinas-zhipu-posts-132-rise-in-annual-revenue-on-ai-boom/","title":"China's Zhipu posts 132% rise in annual revenue on AI boom","publisher":"Reuters (via WHBL)","publishedDate":"2026-03-31","accessedDate":"2026-10-22","tier":"T5","locator":"2025 revenue: on-premise RMB 533.9m; cloud-based revenue from API services RMB 190.4m (+292%); the page does not say whether the GLM models are open or closed"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Inputs for Zhipu AI (Z.ai): Zhipu's flagship GLM models are open-weight (card; Open-weight sublayer); no closed overseas model sales found. Allowance US$0.005bn (0-0.05). Coding fraction 0.40 (0.20-0.60). Re-checked in the full pass: the GLM flagships are open-weight, so their API revenue belongs to Open-weight; no closed Zhipu model was found, so the allowance is kept. General spend = 0.005 × (1 − 0.4) = US$3m (low 0 × 0.4 = 0m; high 0.05 × 0.8 = 40m). Part total US$48.48bn. Share 0.00619% (range 0-0.0824%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-03-31 (full date)."},"china-api":{"low":0,"central":1.32,"high":6.27,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://eu.36kr.com/en/p/3966251451209352","title":"Zhipu AI’s 2B Yuan H1 Loss: Can Selling AI Tokens on Tmall Fix Its Profitability Dilemma?","publisher":"36Kr","publishedDate":"2026-09-03","accessedDate":"2026-10-22","tier":"T5","locator":"ARR of US$1.6bn at the end of August 2026, calculated as August revenue × 12 (above US$2bn on latest week × 52); Coding Plan packages sold online; no mainland-China or open-weight/closed split"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for Zhipu AI (Z.ai): IDC (7 May 2026, https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/, T4) ranks Zhipu fourth by 2025 China public-cloud MaaS revenue; Reuters (31 Mar 2026, T3) reports 2025 cloud API revenue of RMB 190.4m (+292%), about 6% of IDC's RMB 3.07bn market. Run-rate, from pages opened on 22 October 2026: Zhipu's MaaS platform ARR was US$1bn in July 2026 (36Kr, 17 Jul 2026, https://eu.36kr.com/en/p/3898662052693894, T3, from 'multiple independent sources'; Zhipu had not responded) and US$1.6bn at the end of August 2026, calculated as August revenue × 12 (36Kr, 3 Sep 2026, the source above, T3; Investinchina, 31 Aug 2026, https://investinchina.asia/finance/2026-08-z-ai-zhihpu-h1-2026-revenue-surges-399-7-as-business-model-pivots-to-cloud-apis-adjusted-loss-still-widens-443, T3, reporting the interim results); H1 2026 open-platform and API revenue was RMB 825.2m, 86.5% of RMB 953.9m. These figures are for Zhipu's MaaS platform as a whole; the pages give no split between mainland China and overseas, between open-weight and closed models, or by product (such as the Coding Plan). At RMB 7.15 per US dollar the US$1.6bn is RMB 11.44bn, 94% of the RMB 12.11bn part total for closed models in mainland China, so it cannot be mainly closed-model mainland revenue; the GLM flagships are open-weight and count in Open-weight. The earlier input (RMB 1.7bn in 2026, from a report never opened) is dropped. No opened source gives revenue for a closed Zhipu model in mainland China, so the allowance is a 5LayerCake judgement (T5): RMB 0.16bn (0-0.8), about 1.4% (high end 7%) of the RMB 11.44bn ARR. No closed Zhipu model was confirmed, so commercial activity in this sublayer is not marked verified. Company revenue RMB 0.16bn (low 0, high 0.8); part total RMB 12.11bn. Share 1.32% (range 0-6.27%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source published date: 2026-09-03 (full date); the IDC ranking (7 May 2026) is the market anchor."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"d1a3d5081e77ccd738d8a0e617b8eee66c33a9dfd9183ddfbf4f247ada6d491d"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"d1a3d5081e77ccd738d8a0e617b8eee66c33a9dfd9183ddfbf4f247ada6d491d"}}},"huawei-cloud":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"huawei-cloud","name":"Huawei Cloud (Pangu)","kind":"unit","parentId":"huawei","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00149,"high":0.0268,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://futurecio.tech/huawei-cloud-unveils-innovation-and-practices-to-empower-ai-pioneers","title":"Huawei Cloud unveils innovation and practices to empower AI pioneers","publisher":"FutureCIO","publishedDate":"2025-10-01","accessedDate":"2026-10-22","tier":"T5","locator":"Pangu models applied in more than 500 scenarios across 30+ industries (Huawei Cloud keynote, Huawei Connect 2025); no Pangu MaaS revenue or closed/open split"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Huawei Cloud (Pangu): No overseas Pangu sales found (Huawei Cloud's international MaaS lists GLM, DeepSeek and Qwen, which count in Inference serving); allowance US$0.005bn (0-0.03). Coding fraction 0.10 (0-0.30). Unit only, never Huawei's totals. Coding spend = 0.005 × 0.1 = US$0.5m (low 0 × 0 = 0m; high 0.03 × 0.3 = 9m). Part total US$33.54bn. Share 0.00149% (range 0-0.0268%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2025-10-01 (full date)."},"general-reasoning":{"low":0,"central":0.00928,"high":0.0618,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://futurecio.tech/huawei-cloud-unveils-innovation-and-practices-to-empower-ai-pioneers","title":"Huawei Cloud unveils innovation and practices to empower AI pioneers","publisher":"FutureCIO","publishedDate":"2025-10-01","accessedDate":"2026-10-22","tier":"T5","locator":"Pangu models applied in more than 500 scenarios across 30+ industries (Huawei Cloud keynote, Huawei Connect 2025); no Pangu MaaS revenue or closed/open split"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Huawei Cloud (Pangu): No overseas Pangu sales found (Huawei Cloud's international MaaS lists GLM, DeepSeek and Qwen, which count in Inference serving); allowance US$0.005bn (0-0.03). Coding fraction 0.10 (0-0.30). Unit only, never Huawei's totals. General spend = 0.005 × (1 − 0.1) = US$4.5m (low 0 × 0.7 = 0m; high 0.03 × 1 = 30m). Part total US$48.48bn. Share 0.00928% (range 0-0.0618%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2025-10-01 (full date)."},"china-api":{"low":0.25,"central":1.24,"high":4.78,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","title":"中国MaaS市场进入高速增长期，Token经济从概念走向规模化实践 (China's MaaS market enters a high-growth phase)","publisher":"IDC","publishedDate":"2026-05-07","accessedDate":"2026-10-22","tier":"T5","locator":"2025 public-cloud MaaS revenue ranking: Volcano Engine (over 40% share), Alibaba Cloud, Baidu AI Cloud, Zhipu, China Mobile Cloud; Tencent Cloud, SenseTime, Huawei Cloud and China Telecom Cloud named as other vendors"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for Huawei Cloud (Pangu): IDC names Huawei Cloud among MaaS vendors without a share (T4). Huawei says Pangu models are applied in more than 500 scenarios across 30+ industries (Oct 2025, T2) but gives no Pangu API revenue or closed/open split. Calculation: market RMB 22bn × Huawei share 2% (0.5-4%) × Pangu closed 35% (10-70%) = RMB 0.15bn (0.03-0.6). Company revenue RMB 0.15bn (low 0.03, high 0.6); part total RMB 12.11bn. Share 1.24% (range 0.25-4.78%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source published date: 2026-05-07 (full date)."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c912ace5101c536bdfc302a4348c6ba6fe76d27cef996bf7197e22efe3168476"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c912ace5101c536bdfc302a4348c6ba6fe76d27cef996bf7197e22efe3168476"}}},"stepfun":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"stepfun","name":"StepFun","kind":"private","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00179,"high":0.0298,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.thestandard.com.hk/finance/article/331515/Stepfun-Chinas-AI-Six-Tigers-finishes-new-US25b-funding-round-for-HK-IPO","title":"StepFun, China's AI Six Tigers, finishes new US$2.5b funding round for HK IPO","publisher":"The Standard (Hong Kong)","publishedDate":"2026-05-08","accessedDate":"2026-10-22","tier":"T5","locator":"Reportedly completed a US$2.5bn round and dismantled its red-chip structure; no revenue, ARR, API business or open/closed information on the page"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for StepFun: No StepFun API revenue is disclosed (a May 2026 report on its US$2.5bn raise, The Standard, gives none); its open Step releases count in Open-weight. Allowance US$0.002bn (0-0.02). Coding fraction 0.30 (0.10-0.50). Coding spend = 0.002 × 0.3 = US$0.6m (low 0 × 0.1 = 0m; high 0.02 × 0.5 = 10m). Part total US$33.54bn. Share 0.00179% (range 0-0.0298%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-05-08 (full date)."},"general-reasoning":{"low":0,"central":0.00289,"high":0.0371,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.thestandard.com.hk/finance/article/331515/Stepfun-Chinas-AI-Six-Tigers-finishes-new-US25b-funding-round-for-HK-IPO","title":"StepFun, China's AI Six Tigers, finishes new US$2.5b funding round for HK IPO","publisher":"The Standard (Hong Kong)","publishedDate":"2026-05-08","accessedDate":"2026-10-22","tier":"T5","locator":"Reportedly completed a US$2.5bn round and dismantled its red-chip structure; no revenue, ARR, API business or open/closed information on the page"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for StepFun: No StepFun API revenue is disclosed (a May 2026 report on its US$2.5bn raise, The Standard, gives none); its open Step releases count in Open-weight. Allowance US$0.002bn (0-0.02). Coding fraction 0.30 (0.10-0.50). General spend = 0.002 × (1 − 0.3) = US$1.4m (low 0 × 0.5 = 0m; high 0.02 × 0.9 = 18m). Part total US$48.48bn. Share 0.00289% (range 0-0.0371%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-05-08 (full date)."},"china-api":{"low":0,"central":0.661,"high":3.22,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.thestandard.com.hk/finance/article/331515/Stepfun-Chinas-AI-Six-Tigers-finishes-new-US25b-funding-round-for-HK-IPO","title":"StepFun, China's AI Six Tigers, finishes new US$2.5b funding round for HK IPO","publisher":"The Standard (Hong Kong)","publishedDate":"2026-05-08","accessedDate":"2026-10-22","tier":"T5","locator":"Reportedly completed a US$2.5bn round and dismantled its red-chip structure; no revenue, ARR, API business or open/closed information on the page"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for StepFun: No StepFun closed-model API revenue was found; allowance RMB 0.08bn (0-0.4) for closed Step API models. Not marked verified. Company revenue RMB 0.08bn (low 0, high 0.4); part total RMB 12.11bn. Share 0.661% (range 0-3.22%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-05-08 (full date); the IDC ranking is the market anchor."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"fd8ce5f971189e165d85e4e2574224b29a1a090ac0d8f0e36904b0ecfa3fa822"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"fd8ce5f971189e165d85e4e2574224b29a1a090ac0d8f0e36904b0ecfa3fa822"}}},"kuaishou-kwaipilot":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"kuaishou-kwaipilot","name":"Kuaishou Kwaipilot (KAT-Coder)","kind":"unit","parentId":"kuaishou","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0.00209,"central":0.0134,"high":0.0894,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://openrouter.ai/kwaipilot/kat-coder-pro-v2.5/api","title":"KAT-Coder-Pro V2.5 (kwaipilot)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"KAT-Coder-Pro V2.5 listed from 10 July 2026, 256K context; the page shows no price, usage or closed/open status; the page shows no publication date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Kuaishou Kwaipilot (KAT-Coder): KAT-Coder-Pro V2.5 has been listed on OpenRouter since 10 July 2026 (the page shows no price or closed/open status); no revenue is disclosed. Allowance US$0.005bn (0.001-0.03). Coding fraction 0.90 (0.70-1.0): a coding model. Unit only, never Kuaishou's totals. Coding spend = 0.005 × 0.9 = US$4.5m (low 0.001 × 0.7 = 0.7m; high 0.03 × 1 = 30m). Part total US$33.54bn. Share 0.0134% (range 0.00209-0.0894%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: undated (the page shows none)."},"general-reasoning":{"low":0,"central":0.00103,"high":0.0186,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://openrouter.ai/kwaipilot/kat-coder-pro-v2.5/api","title":"KAT-Coder-Pro V2.5 (kwaipilot)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"KAT-Coder-Pro V2.5 listed from 10 July 2026, 256K context; the page shows no price, usage or closed/open status; the page shows no publication date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Kuaishou Kwaipilot (KAT-Coder): KAT-Coder-Pro V2.5 has been listed on OpenRouter since 10 July 2026 (the page shows no price or closed/open status); no revenue is disclosed. Allowance US$0.005bn (0.001-0.03). Coding fraction 0.90 (0.70-1.0): a coding model. Unit only, never Kuaishou's totals. General spend = 0.005 × (1 − 0.9) = US$0.5m (low 0.001 × 0 = 0m; high 0.03 × 0.3 = 9m). Part total US$48.48bn. Share 0.00103% (range 0-0.0186%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: undated (the page shows none)."},"china-api":{"low":0.0828,"central":0.413,"high":1.63,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://openrouter.ai/kwaipilot/kat-coder-pro-v2.5/api","title":"KAT-Coder-Pro V2.5 (kwaipilot)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"KAT-Coder-Pro V2.5 listed from 10 July 2026, 256K context; the page shows no price, usage or closed/open status; the page shows no publication date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for Kuaishou Kwaipilot (KAT-Coder): KAT-Coder models are sold through StreamLake and other platforms (the OpenRouter page confirms a listing only). Assumed mainland API revenue RMB 0.05bn (0.01-0.2). Not marked verified. Company revenue RMB 0.05bn (low 0.01, high 0.2); part total RMB 12.11bn. Share 0.413% (range 0.0828-1.63%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: undated (the page shows none); the IDC ranking is the market anchor."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c1fd0359782b9600986792dc64b7871d4a66ab1bfa30cc285ed67b1881d34dfa"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c1fd0359782b9600986792dc64b7871d4a66ab1bfa30cc285ed67b1881d34dfa"}}},"writer":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"writer","name":"Writer","kind":"private","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.000596,"high":0.0119,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://dev.writer.com/home/models","title":"Writer API: models","publisher":"Writer","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Palmyra X6 $2.00/$8.00, X5 $0.60/$6.00, X4 $2.50/$10.00 per million tokens; X5 and X4 on Amazon Bedrock; X5 and X4 deprecated 14 Dec 2026; the page shows no date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Writer: Writer's API sells Palmyra X6, X5 and X4 per token and X5/X4 on Amazon Bedrock (T2). Writer's revenue is not disclosed and its platform seats count in Enterprise agents; allowance for API spend US$0.01bn (0.002-0.04). Coding fraction 0.02 (0-0.10). Coding spend = 0.01 × 0.02 = US$0.2m (low 0.002 × 0 = 0m; high 0.04 × 0.1 = 4m). Part total US$33.54bn. Share 0.000596% (range 0-0.0119%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: undated (the page shows none)."},"general-reasoning":{"low":0.00371,"central":0.0202,"high":0.0825,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://dev.writer.com/home/models","title":"Writer API: models","publisher":"Writer","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Palmyra X6 $2.00/$8.00, X5 $0.60/$6.00, X4 $2.50/$10.00 per million tokens; X5 and X4 on Amazon Bedrock; X5 and X4 deprecated 14 Dec 2026; the page shows no date"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Writer: Writer's API sells Palmyra X6, X5 and X4 per token and X5/X4 on Amazon Bedrock (T2). Writer's revenue is not disclosed and its platform seats count in Enterprise agents; allowance for API spend US$0.01bn (0.002-0.04). Coding fraction 0.02 (0-0.10). General spend = 0.01 × (1 − 0.02) = US$9.8m (low 0.002 × 0.9 = 1.8m; high 0.04 × 1 = 40m). Part total US$48.48bn. Share 0.0202% (range 0.00371-0.0825%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: undated (the page shows none)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://dev.writer.com/home/models","title":"Writer API: models","publisher":"Writer","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Palmyra X6 $2.00/$8.00, X5 $0.60/$6.00, X4 $2.50/$10.00 per million tokens; X5 and X4 on Amazon Bedrock; X5 and X4 deprecated 14 Dec 2026; the page shows no date"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for Writer's closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: undated (the page shows none); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"a065a9ab0ba229133d7e799c479b0aea46471f8592a88901c4eeb23f96915611"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"a065a9ab0ba229133d7e799c479b0aea46471f8592a88901c4eeb23f96915611"}}},"ntt":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"ntt","name":"NTT (tsuzumi)","kind":"public","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00149,"high":0.0179,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://group.ntt/en/newsrelease/2026/05/19/260519a.html","title":"tsuzumi 2 update","publisher":"NTT","publishedDate":"2026-05-19","accessedDate":"2026-10-22","tier":"T5","locator":"tsuzumi 2 mainly deployed in on-premises and private-cloud environments; gradually made available through NTT Group companies; no public API or revenue mentioned"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for NTT (tsuzumi): NTT says tsuzumi 2 is mainly deployed in on-premises and private-cloud environments (19 May 2026, T2); no public API or revenue is mentioned, and on-premises licences are not API spending. Allowance US$0.01bn (0-0.03). Coding fraction 0.05 (0-0.20). Tsuzumi only, never NTT's totals. Coding spend = 0.01 × 0.05 = US$0.5m (low 0 × 0 = 0m; high 0.03 × 0.2 = 6m). Part total US$33.54bn. Share 0.00149% (range 0-0.0179%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-05-19 (full date)."},"general-reasoning":{"low":0,"central":0.0196,"high":0.0619,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://group.ntt/en/newsrelease/2026/05/19/260519a.html","title":"tsuzumi 2 update","publisher":"NTT","publishedDate":"2026-05-19","accessedDate":"2026-10-22","tier":"T5","locator":"tsuzumi 2 mainly deployed in on-premises and private-cloud environments; gradually made available through NTT Group companies; no public API or revenue mentioned"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for NTT (tsuzumi): NTT says tsuzumi 2 is mainly deployed in on-premises and private-cloud environments (19 May 2026, T2); no public API or revenue is mentioned, and on-premises licences are not API spending. Allowance US$0.01bn (0-0.03). Coding fraction 0.05 (0-0.20). Tsuzumi only, never NTT's totals. General spend = 0.01 × (1 − 0.05) = US$9.5m (low 0 × 0.8 = 0m; high 0.03 × 1 = 30m). Part total US$48.48bn. Share 0.0196% (range 0-0.0619%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-05-19 (full date)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://group.ntt/en/newsrelease/2026/05/19/260519a.html","title":"tsuzumi 2 update","publisher":"NTT","publishedDate":"2026-05-19","accessedDate":"2026-10-22","tier":"T5","locator":"tsuzumi 2 mainly deployed in on-premises and private-cloud environments; gradually made available through NTT Group companies; no public API or revenue mentioned"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for NTT (tsuzumi)'s closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-05-19 (full date); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"b9cdfbfc57ab14d37a370194e2502d6b844fbaca0709d9d2ef3b845f45c0958b"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"b9cdfbfc57ab14d37a370194e2502d6b844fbaca0709d9d2ef3b845f45c0958b"}}},"upstage":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"upstage","name":"Upstage","kind":"private","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00149,"high":0.0298,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://console.upstage.ai/docs/getting-started/models","title":"Upstage Console: models","publisher":"Upstage","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Solar Mini 4 listed as live with a 50% promotion through 22 October; no base prices, revenue or open/closed status on the page; the page shows no year"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Upstage: Upstage's console lists Solar Mini 4 as live with a 50% promotion through 22 October (T2); revenue is not disclosed. Allowance US$0.01bn (0.003-0.05). Coding fraction 0.05 (0-0.20). Coding spend = 0.01 × 0.05 = US$0.5m (low 0.003 × 0 = 0m; high 0.05 × 0.2 = 10m). Part total US$33.54bn. Share 0.00149% (range 0-0.0298%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: undated (the page shows none)."},"general-reasoning":{"low":0.00495,"central":0.0196,"high":0.103,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://console.upstage.ai/docs/getting-started/models","title":"Upstage Console: models","publisher":"Upstage","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Solar Mini 4 listed as live with a 50% promotion through 22 October; no base prices, revenue or open/closed status on the page; the page shows no year"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Upstage: Upstage's console lists Solar Mini 4 as live with a 50% promotion through 22 October (T2); revenue is not disclosed. Allowance US$0.01bn (0.003-0.05). Coding fraction 0.05 (0-0.20). General spend = 0.01 × (1 − 0.05) = US$9.5m (low 0.003 × 0.8 = 2.4m; high 0.05 × 1 = 50m). Part total US$48.48bn. Share 0.0196% (range 0.00495-0.103%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: undated (the page shows none)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://console.upstage.ai/docs/getting-started/models","title":"Upstage Console: models","publisher":"Upstage","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Solar Mini 4 listed as live with a 50% promotion through 22 October; no base prices, revenue or open/closed status on the page; the page shows no year"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for Upstage's closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: undated (the page shows none); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"1508eee1ac1315d70f9089f9e3f52070f1196e8d3de1f379617889f21af9b1f9"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"1508eee1ac1315d70f9089f9e3f52070f1196e8d3de1f379617889f21af9b1f9"}}},"minimax":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"minimax","name":"MiniMax","kind":"public","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0,"central":0.00596,"high":0.0536,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results","title":"MiniMax Global Announces Full Year 2025 Financial Results","publisher":"MiniMax","publishedDate":"2026-03-02","accessedDate":"2026-10-22","tier":"T5","locator":"Open Platform and enterprise revenue US$26.0m in 2025 (+197.8%); total revenue US$79.0m, over 70% from overseas; no open/closed split of models"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for MiniMax: MiniMax's Open Platform and enterprise revenue was US$26.0m in 2025 with over 70% of total revenue from overseas (company results, 2 Mar 2026, T2); its M2.x text models are open-weight (register) and speech, music and video count in Specialist models. Closed text-model allowance US$0.005bn (0-0.03). Coding fraction 0.40 (0.20-0.60). Coding spend = 0.005 × 0.4 = US$2m (low 0 × 0.2 = 0m; high 0.03 × 0.6 = 18m). Part total US$33.54bn. Share 0.00596% (range 0-0.0536%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2026-03-02 (full date)."},"general-reasoning":{"low":0,"central":0.00619,"high":0.0495,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results","title":"MiniMax Global Announces Full Year 2025 Financial Results","publisher":"MiniMax","publishedDate":"2026-03-02","accessedDate":"2026-10-22","tier":"T5","locator":"Open Platform and enterprise revenue US$26.0m in 2025 (+197.8%); total revenue US$79.0m, over 70% from overseas; no open/closed split of models"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for MiniMax: MiniMax's Open Platform and enterprise revenue was US$26.0m in 2025 with over 70% of total revenue from overseas (company results, 2 Mar 2026, T2); its M2.x text models are open-weight (register) and speech, music and video count in Specialist models. Closed text-model allowance US$0.005bn (0-0.03). Coding fraction 0.40 (0.20-0.60). General spend = 0.005 × (1 − 0.4) = US$3m (low 0 × 0.4 = 0m; high 0.03 × 0.8 = 24m). Part total US$48.48bn. Share 0.00619% (range 0-0.0495%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2026-03-02 (full date)."},"china-api":{"low":0,"central":0.248,"high":1.63,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results","title":"MiniMax Global Announces Full Year 2025 Financial Results","publisher":"MiniMax","publishedDate":"2026-03-02","accessedDate":"2026-10-22","tier":"T5","locator":"Open Platform and enterprise revenue US$26.0m in 2025 (+197.8%); total revenue US$79.0m, over 70% from overseas; no open/closed split of models"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Method: share = company closed-model API revenue in mainland China ÷ part total (the sum of all companies' estimates, so shares cannot exceed 100%); low and high move this company's revenue to its range ends with the others at central. Inputs for MiniMax: MiniMax's whole Open Platform and enterprise revenue was US$26.0m (RMB 0.19bn) in 2025, with over 70% of total revenue from overseas (T2). Closed text-model API in mainland China is assumed 15% (0-100%) of the domestic part: RMB 0.03bn (0-0.2). Not marked verified because no closed text model was confirmed. Company revenue RMB 0.03bn (low 0, high 0.2); part total RMB 12.11bn. Share 0.248% (range 0-1.63%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2026-03-02 (full date); the IDC ranking is the market anchor."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"77fb85abefdebb32e95d84d35b95979e65bca2ac041317491533e634024325f2"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"77fb85abefdebb32e95d84d35b95979e65bca2ac041317491533e634024325f2"}}},"inception-labs":{"schema":"5lc-index-inputs/1","key":"models-frontier-labs","entityId":"inception-labs","name":"Inception","kind":"private","commercialActivityVerified":false,"shares":{"coding-agents":{"low":0.000298,"central":0.00358,"high":0.0179,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://the-decoder.com/inceptions-comeback-brings-new-funding-and-a-shift-to-diffusion-models","title":"Inception's comeback brings new funding and a shift to diffusion models","publisher":"The Decoder","publishedDate":"2025-11-09","accessedDate":"2026-10-22","tier":"T5","locator":"Mercury available through OpenRouter and Poe at $0.25/$1 per million tokens; Menlo-led raise; no customers named, no revenue or usage"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Inception: Mercury diffusion models are sold through OpenRouter and Poe at $0.25/$1 per million tokens (The Decoder, 9 Nov 2025, T3, after a US$50m raise); no customers named and no revenue or usage disclosed. Allowance US$0.003bn (0.0005-0.01). Coding fraction 0.40 (0.20-0.60): Mercury Coder. Coding spend = 0.003 × 0.4 = US$1.2m (low 0.0005 × 0.2 = 0.1m; high 0.01 × 0.6 = 6m). Part total US$33.54bn. Share 0.00358% (range 0.000298-0.0179%). Part total, coding spend outside mainland China, annualised US$bn (central): anthropic 25.9, openai 4.56, google-deepmind 2.16, xai 0.24, amazon-agi 0.06, mistral 0.05, meta-superintelligence-labs 0.036, alibaba-qwen 0.0315, bytedance-seed 0.03, microsoft-ai 0.012, moonshot 0.008, tencent-hy 0.006, sber 0.005, smaller register companies 0.022, others outside the register 0.4; central total 33.54 (about 41% of the US$ 82.0bn spend; the card's weight is 34.8% with range 25.5-54.5%). Source publication date: 2025-11-09 (full date)."},"general-reasoning":{"low":0.000413,"central":0.00371,"high":0.0165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://the-decoder.com/inceptions-comeback-brings-new-funding-and-a-shift-to-diffusion-models","title":"Inception's comeback brings new funding and a shift to diffusion models","publisher":"The Decoder","publishedDate":"2025-11-09","accessedDate":"2026-10-22","tier":"T5","locator":"Mercury available through OpenRouter and Poe at $0.25/$1 per million tokens; Menlo-led raise; no customers named, no revenue or usage"},"note":"5LayerCake estimate (T5), full-inputs pass of 22 October 2026. Method: share = company spend in the part ÷ part total, where the part total is the sum of all developers' estimates in the part, so the shares cannot exceed 100%. Company spend in the part = annualised closed-model API spend outside mainland China, counted for the developer across all channels, × coding fraction (coding) or × (1 − coding fraction) (general). Low and high move this company's spend and coding fraction to their range ends while the other companies stay at central. Light upgrade (efficiency rule): the high-end screening Importance is below the tenth-place central estimate, so the screening estimate is kept and only the part totals were refreshed. Inputs for Inception: Mercury diffusion models are sold through OpenRouter and Poe at $0.25/$1 per million tokens (The Decoder, 9 Nov 2025, T3, after a US$50m raise); no customers named and no revenue or usage disclosed. Allowance US$0.003bn (0.0005-0.01). Coding fraction 0.40 (0.20-0.60): Mercury Coder. General spend = 0.003 × (1 − 0.4) = US$1.8m (low 0.0005 × 0.4 = 0.2m; high 0.01 × 0.8 = 8m). Part total US$48.48bn. Share 0.00371% (range 0.000413-0.0165%). Part total, general spend outside mainland China, annualised US$bn (central): anthropic 28.1, openai 11.7, google-deepmind 5.84, xai 0.56, amazon-agi 0.34, mistral 0.15, meta-superintelligence-labs 0.084, cohere 0.0784, bytedance-seed 0.07, alibaba-qwen 0.0585, smaller register companies 0.29, others outside the register 1.2; central total 48.48. Source publication date: 2025-11-09 (full date)."},"china-api":{"low":0,"central":0.0248,"high":0.165,"unit":"percent","basis":"Share of the function's annualised run-rate of model API spending in US dollars, counted for the model's developer across all sales channels (latest month × 12 or latest quarter × 4)","period":"Q3 2026, annualised","periodEnd":"2026-09-30","source":{"url":"https://the-decoder.com/inceptions-comeback-brings-new-funding-and-a-shift-to-diffusion-models","title":"Inception's comeback brings new funding and a shift to diffusion models","publisher":"The Decoder","publishedDate":"2025-11-09","accessedDate":"2026-10-22","tier":"T5","locator":"Mercury available through OpenRouter and Poe at $0.25/$1 per million tokens; Menlo-led raise; no customers named, no revenue or usage"},"note":"5LayerCake allowance (T5), full-inputs pass of 22 October 2026: no mainland-China sales channel for Inception's closed models was found in the sources opened. Allowance RMB 0.003bn (0-0.02) a year, the same for every foreign developer without a found channel, as a share of the part total (RMB 12.11bn). Share 0.0248% (range 0-0.165%). Part total, closed-model API revenue in mainland China, annualised RMB bn (central): bytedance-seed 5.3, alibaba-qwen 3.7, baidu-ernie 0.9, tencent-hy 0.4, iflytek 0.4, sensetime 0.25, china-mobile 0.2, zhipu 0.16, huawei-cloud 0.15, moonshot 0.15, smaller register companies 0.2, others outside the register 0.3; central total 12.11 (US$ 1.69bn at RMB 7.15; the card's total is US$1.69bn, range 1.04-2.74). Source publication date: 2025-11-09 (full date); the source describes the company's products, not its China sales."}},"screening":{"coding-agents":{"levelLow":1,"levelHigh":2},"general-reasoning":{"levelLow":1,"levelHigh":2},"china-api":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c78b19d2ee524d6c29f42090e96f5cea3497d718a2add595cc223395d254f5e1"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c78b19d2ee524d6c29f42090e96f5cea3497d718a2add595cc223395d254f5e1"}}}},"ratings":{"anthropic":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"anthropic","rater":"a","model":"claude-opus-5-5","dachiReview":true,"ratedAt":"2026-10-22","functions":{"coding-agents":{"alternatives":{"level":3,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra","date":"2026-10-09","support":"Artificial Analysis Coding Agent Index: Claude Sonnet 5.5 and Opus 5.5 top two; GPT-6.1 Sol in Codex next (full date)."},{"url":"https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/","date":"2025-12-09","support":"Menlo (an Anthropic investor) estimates coding share Anthropic 54%, OpenAI 21%; all API spend 40/27/21% with Google (full date)."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Anthropic's gateway spend share fell from 63.7% to 43.9% in September 2026 as GPT-6 and open-weight models gained (full date)."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Vercel says nearly one in two gateway teams changed their most-used model in September 2026, almost one in three their lab (full date)."},{"url":"https://fortune.com/2026/09/11/openai-astra-chatgpt-pro-pause/","date":"2026-09-11","support":"OpenAI said it lacked capacity for GPT-6 Astra demand, limited first enterprise access and paused Pro sign-ups (full date)."},{"url":"https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/","date":"2026-07-22","support":"The company says its model APIs process about 22bn tokens a minute and 'we continue to be supply constrained' (full date)."}]},"capabilityGap":{"level":3,"evidence":[{"url":"https://www.anthropic.com/news/claude-opus-5-5","date":"2026-09-22","support":"The company says Opus 5.5 scores 66.4% on Terminal-Bench 4.0 against 57.9% for GPT-6 Astra (benchmark table; full date)."},{"url":"https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra","date":"2026-10-09","support":"Artificial Analysis Coding Agent Index: Claude Sonnet 5.5 and Opus 5.5 top two; GPT-6.1 Sol in Codex next (full date)."},{"url":"https://fortune.com/2026/09/11/openai-astra-chatgpt-pro-pause/","date":"2026-09-11","support":"OpenAI said it lacked capacity for GPT-6 Astra demand, limited first enterprise access and paused Pro sign-ups (full date)."},{"url":"https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/","date":"2026-07-22","support":"Google, the next-largest alternative, says it remains supply constrained; replacing a three-quarter share needs large new capacity (full date)."}]}},"general-reasoning":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Intelligence Index: Claude Opus 5.5 58, Meta Muse Wonder Full 56, GPT-6 Astra 53, Gemini 4 Argon 53; page shows no date."},{"url":"https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/","date":"2025-12-09","support":"Menlo (an Anthropic investor) estimates coding share Anthropic 54%, OpenAI 21%; all API spend 40/27/21% with Google (full date)."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Open-weight models were over 70% of September gateway tokens and nearly a quarter of spend (full date)."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Vercel says nearly one in two gateway teams changed their most-used model in September 2026, almost one in three their lab (full date)."},{"url":"https://aws.amazon.com/about-aws/whats-new/2026/09/openai-gpt-6-astra-on-amazon-bedrock/","date":"2026-09-08","support":"GPT-6 Astra is generally available on Amazon Bedrock, and Codex can be set up to run on Bedrock (full date)."},{"url":"https://fortune.com/2026/09/11/openai-astra-chatgpt-pro-pause/","date":"2026-09-11","support":"OpenAI said it lacked capacity for GPT-6 Astra demand, limited first enterprise access and paused Pro sign-ups (full date)."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Intelligence Index: Claude Opus 5.5 58, Meta Muse Wonder Full 56, GPT-6 Astra 53, Gemini 4 Argon 53; page shows no date."},{"url":"https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/","date":"2026-07-22","support":"The company says its model APIs process about 22bn tokens a minute and 'we continue to be supply constrained' (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"anthropic","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","dachiReview":true,"functions":{"coding-agents":{"alternatives":{"level":3,"evidence":[{"url":"https://datacamp.com/blog/best-llm-for-coding","date":"2026-09-25","support":"A 25 Sep 2026 comparison table puts OpenAI's GPT-6 Astra level with Opus 5.5 on Terminal-Bench 4.0 (59.6% each on Artificial Analysis's run); other rows are Google Gemini 3.8 Flash and open-weight DeepSeek and Kimi. Full date shown on the page."},{"url":"https://vercel.com/blog/ai-gateway-production-index-september-2026","date":"2026-09-17","support":"GPT-6 Astra took 7.7% of all gateway spend in its first twelve days (Fable 5.1: 3.7%), so OpenAI is a customer-proven alternative to Anthropic's flagship. Full date shown on the page."},{"url":"https://ai.google.dev/gemini-api/docs/changelog","date":"2026-10-13","support":"Google made gemini-4-argon, which it calls its most capable model for complex coding and agents, generally available on 13 Oct 2026, nine days before this rating: an announced-but-barely-qualified entrant. Full date shown on the page."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two teams changed their most-used model and almost one in three changed lab: customers re-point within weeks. The limit on replacing Anthropic's volume is alternatives' capacity, not customer qualification (rater inference). Full date shown on the page."},{"url":"https://thenextweb.com/news/anthropic-q3-revenue-16-8bn-ipo","date":"2026-10-15","support":"Reports of Anthropic's own quarterly revenue ($4.7bn in Q1, $11.6bn in Q2, $16.8bn in Q3 2026) show a 3.6x scale-up in two quarters, which suggests an alternative can add this scale in under a year (rater inference; figures first reported by the FT, not published by Anthropic). Full date shown on the page."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://www.anthropic.com/news/claude-opus-5-5","date":"2026-09-22","support":"The company says Opus 5.5 scores 66.4% against GPT-6 Astra's 57.9% on Terminal-Bench 4.0 and 54.4% against 53.3% on FrontierCode v1.1 (company's own table; a lead of 1 to 8.5 points). Full date shown on the page."},{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Artificial Analysis Intelligence Index lists Claude Opus 5.5 at 58 and GPT-6 Astra and Gemini 4 Argon at 53 each; Grok 4.7 at 46. Live page with no publication date (dated 'undated')."}]}},"general-reasoning":{"alternatives":{"level":2,"evidence":[{"url":"https://openrouter.ai/rankings","date":"undated","support":"OpenRouter's capability ranking (Artificial Analysis scores, usage data through 22 Oct 2026) lists Claude Opus 5.5 57.6, Qwen3.8 Max 53.4, GPT-6 Astra 52.7, Gemini 4 Argon 52.6, Muse Spark 1.3 48.1: three other labs within about 5 points and Meta a fourth. Live page with no publication date (dated 'undated')."},{"url":"https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder","date":"2026-10-12","support":"Meta released Muse Wonder through its Model API on 12 Oct 2026 at $4/$20 per million tokens, priced like Opus 5.5 and GPT-6 Sol: a fifth entrant, but days old. Full date shown on the page."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two teams changed their most-used model, and Anthropic's gateway spend share fell from 63.7% to 43.9% in one month: customers move in weeks. Absorbing more than half of general-purpose API spend still needs months of added capacity at the others (rater inference). Full date shown on the page."},{"url":"https://www.khaleejtimes.com/business/anthropic-ipo-prospectus-lays-bare-deep-dependence-on-big-tech-partners","date":"2026-09-30","support":"Reuters on Anthropic's IPO filing: long-term compute commitments above $417bn covering 3.5 GW, and many of its biggest customers have no long-term contracts. Full date shown on the page."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://www.anthropic.com/news/claude-opus-5-5","date":"2026-09-22","support":"The company's table shows Opus 5.5 ahead of GPT-6 Astra on GDPval-AA (1846 v 1542 Elo) and Humanity's Last Exam with tools (67.7% v 57.2%), behind on Terminal-Bench-Science and AutomationBench (company's own claims). Full date shown on the page."},{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Artificial Analysis shows Opus 5.5 at 58 against 53 for GPT-6 Astra and Gemini 4 Argon: a small lead over the best alternatives. Live page with no publication date (dated 'undated')."}]}}}}},"openai":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"openai","rater":"a","model":"claude-opus-5-5","dachiReview":true,"ratedAt":"2026-10-22","functions":{"coding-agents":{"alternatives":{"level":2,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra","date":"2026-10-09","support":"Artificial Analysis Coding Agent Index: Claude Sonnet 5.5 and Opus 5.5 top two; GPT-6.1 Sol in Codex next (full date)."},{"url":"https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/","date":"2025-12-09","support":"Menlo (an Anthropic investor) estimates coding share Anthropic 54%, OpenAI 21%; all API spend 40/27/21% with Google (full date)."},{"url":"https://github.blog/changelog/2026-09-03-gemini-3-8-flash-is-now-available-in-github-copilot","date":"2026-09-03","support":"Gemini 3.8 Flash added to GitHub Copilot's model picker; strong on terminal coding tasks in early testing (full date)."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Vercel says nearly one in two gateway teams changed their most-used model in September 2026, almost one in three their lab (full date)."},{"url":"https://www.anthropic.com/news/claude-in-microsoft-foundry","date":"2025-11-18","support":"The company says Claude models are offered in Microsoft Foundry to Azure customers, the channel that also sells OpenAI models (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra","date":"2026-10-09","support":"Artificial Analysis Coding Agent Index: Claude Sonnet 5.5 and Opus 5.5 top two; GPT-6.1 Sol in Codex next (full date)."},{"url":"https://www.anthropic.com/news/claude-opus-5-5","date":"2026-09-22","support":"The company says Opus 5.5 scores 66.4% on Terminal-Bench 4.0 against 57.9% for GPT-6 Astra (benchmark table; full date)."},{"url":"https://www.theregister.com/ai-and-ml/2026/05/06/claude-hitches-a-ride-on-spacexs-datacenter-capacity/5231252","date":"2026-05-06","support":"Anthropic's capacity constraints had left Claude customers without service until it took all of Colossus 1, over 300 MW (full date)."}]}},"general-reasoning":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Intelligence Index: Claude Opus 5.5 58, Meta Muse Wonder Full 56, GPT-6 Astra 53, Gemini 4 Argon 53; page shows no date."},{"url":"https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/","date":"2025-12-09","support":"Menlo (an Anthropic investor) estimates coding share Anthropic 54%, OpenAI 21%; all API spend 40/27/21% with Google (full date)."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Launch-window spend shares: GPT-6 Astra 4.86%, Gemini 3.8 Flash 2.21%, Grok 4.7 2.00% (full date)."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://www.anthropic.com/news/claude-in-microsoft-foundry","date":"2025-11-18","support":"The company says Claude models are offered in Microsoft Foundry to Azure customers, the channel that also sells OpenAI models (full date)."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Vercel says nearly one in two gateway teams changed their most-used model in September 2026, almost one in three their lab (full date)."},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"'Switching between vendors is relatively easy, but increasingly rare'; 11% of teams switched vendors in a year (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra","date":"2026-10-09","support":"Artificial Analysis Intelligence Index: Claude Opus 5.5 58, Sonnet 5.5 56, GPT-6 Astra 53, GPT-6.1 Sol 52 (full date)."},{"url":"https://www.theregister.com/ai-and-ml/2026/05/06/claude-hitches-a-ride-on-spacexs-datacenter-capacity/5231252","date":"2026-05-06","support":"Anthropic's capacity constraints had left Claude customers without service until it took all of Colossus 1, over 300 MW (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"openai","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","dachiReview":true,"functions":{"coding-agents":{"alternatives":{"level":2,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Claude Opus 5.5 took 11.03% of gateway spend in its first two weeks; Gemini 3.8 Flash 2.21% and Grok 4.7 2.00% of September spend; open-weight models about 23% of spend: several alternatives are already in production use. Full date shown on the page."},{"url":"https://www.anthropic.com/news/claude-opus-5-5","date":"2026-09-22","support":"The company says Opus 5.5 is available on Amazon Web Services, Google Cloud and Microsoft Azure, so customers buying OpenAI models through Azure can reach a rival coding model on the same platform. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-september-2026","date":"2026-09-17","support":"Astra took about a third of OpenAI's own spend within two days of launch, and nine in ten teams that ran Fable cut its usage in August: coding spend re-allocates in days to weeks. OpenAI's coding volume is small next to Anthropic's (inputs file: 13.6% v 77.3%; rater inference). Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.anthropic.com/news/claude-opus-5-5","date":"2026-09-22","support":"The company's table has Opus 5.5 ahead of GPT-6 Astra on Terminal-Bench 4.0 (66.4% v 57.9%) and FrontierCode v1.1 (54.4% v 53.3%): the best alternative leads in coding (company's claim). Full date shown on the page."},{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Artificial Analysis lists Claude Opus 5.5 at 58 against GPT-6 Astra at 53. Live page with no publication date (dated 'undated')."}]}},"general-reasoning":{"alternatives":{"level":2,"evidence":[{"url":"https://openrouter.ai/rankings","date":"undated","support":"OpenRouter's Artificial Analysis ranking has Claude Opus 5.5 57.6, Sonnet 5.5 56.0, Fable 5.1 53.4 and Qwen3.8 Max 53.4 ahead of GPT-6 Astra at 52.7, with Gemini 4 Argon at 52.6: four other labs at or above Astra. Live page with no publication date (dated 'undated')."},{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Artificial Analysis also lists Meta's Muse Wonder Full at 56 and Gemini 4 Argon at 53, both within a few points of Astra. Live page with no publication date (dated 'undated')."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Nearly one in two teams changed their most-used model in September and Opus 5.5 took 11.03% of spend in two weeks: customers re-point in weeks. OpenAI's general-purpose volume is about a quarter of the market, so alternatives must add capacity, which takes months (rater inference). Full date shown on the page."},{"url":"https://simonwillison.net/2026/jul/12/bump","date":"2026-07-12","support":"Simon Willison (12 Jul 2026): Anthropic limited Fable 5 access citing compute constraints, while OpenAI seemed confident it would not need to restrict GPT-5.6; the strongest rival was capacity-limited then. Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Claude Opus 5.5 (58), Sonnet 5.5 (56) and Fable 5.1 (53) rank at or above GPT-6 Astra (53) on Artificial Analysis: the best alternative is ahead in capability. Live page with no publication date (dated 'undated')."},{"url":"https://www.anthropic.com/news/claude-opus-5-5","date":"2026-09-22","support":"The company's table shows Astra ahead of Opus 5.5 only on AutomationBench (41.4% v 40.0%) and Terminal-Bench-Science (64.6% v 58.7%): narrow pockets (company's own claims). Full date shown on the page."}]}}}}},"google-deepmind":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"google-deepmind","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"coding-agents":{"alternatives":{"level":2,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra","date":"2026-10-09","support":"Artificial Analysis Coding Agent Index: Claude Sonnet 5.5 and Opus 5.5 top two; GPT-6.1 Sol in Codex next (full date)."},{"url":"https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/","date":"2025-12-09","support":"Menlo (an Anthropic investor) estimates coding share Anthropic 54%, OpenAI 21%; all API spend 40/27/21% with Google (full date)."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Open-weight models were over 70% of September gateway tokens and nearly a quarter of spend (full date)."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/opus-5","date":"2026-10-22","support":"Claude Opus 5 is generally available on Google Cloud's agent platform beside Gemini (page 'last updated' date, full)."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Vercel says nearly one in two gateway teams changed their most-used model in September 2026, almost one in three their lab (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra","date":"2026-10-09","support":"Artificial Analysis Coding Agent Index: Claude Sonnet 5.5 and Opus 5.5 top two; GPT-6.1 Sol in Codex next (full date)."},{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Intelligence Index: Claude Opus 5.5 58, Meta Muse Wonder Full 56, GPT-6 Astra 53, Gemini 4 Argon 53; page shows no date."}]}},"general-reasoning":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Intelligence Index: Claude Opus 5.5 58, Meta Muse Wonder Full 56, GPT-6 Astra 53, Gemini 4 Argon 53; page shows no date."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Launch-window spend shares: GPT-6 Astra 4.86%, Gemini 3.8 Flash 2.21%, Grok 4.7 2.00% (full date)."},{"url":"https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/","date":"2025-12-09","support":"Enterprise open-source share fell to 11%; enterprises 'remain cautious, preferring closed-source models' (full date)."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/opus-5","date":"2026-10-22","support":"Claude Opus 5 is generally available on Google Cloud's agent platform beside Gemini (page 'last updated' date, full)."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Vercel says nearly one in two gateway teams changed their most-used model in September 2026, almost one in three their lab (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Intelligence Index: Claude Opus 5.5 58, Meta Muse Wonder Full 56, GPT-6 Astra 53, Gemini 4 Argon 53; page shows no date."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Open-weight models were over 70% of September gateway tokens and nearly a quarter of spend (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"google-deepmind","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"coding-agents":{"alternatives":{"level":1,"evidence":[{"url":"https://datacamp.com/blog/best-llm-for-coding","date":"2026-09-25","support":"Sep 2026 table: Gemini 3.8 Flash 89.4% on Terminal-Bench 2.1 (vendor) against DeepSeek V4.1 Flash 90.6% and Kimi K3 88.3%; Anthropic's Fable 5.1 91.4%; Qwen3.8-27B 73.0%: many labs sit in the same coding band. Full date shown on the page."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Open-weight models processed over 70% of gateway tokens and about 23% of spend in September; Gemini 3.8 Flash held 2.21% of spend: customers already run many qualified alternatives. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-september-2026","date":"2026-09-17","support":"Gemini 3 Flash lost 95% of its token share since May and more than three-quarters of the lost volume went to other labs: Google's volume has already moved to rivals within months. Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Artificial Analysis lists Claude Opus 5.5 (58) and Sonnet 5.5 (56) above Gemini 4 Argon (53); no alternative is short of Google's capability. Live page with no publication date (dated 'undated')."},{"url":"https://datacamp.com/blog/best-llm-for-coding","date":"2026-09-25","support":"Opus 5.5 and GPT-6 Astra tie at 59.6% on Terminal-Bench 4.0 on Artificial Analysis's run, and Fable 5.1 leads Terminal-Bench 2.1 at 91.4%: Google's models are not ahead of their rivals in coding. Full date shown on the page."}]}},"general-reasoning":{"alternatives":{"level":1,"evidence":[{"url":"https://openrouter.ai/rankings","date":"undated","support":"OpenRouter's Artificial Analysis ranking lists Opus 5.5 57.6, Sonnet 5.5 56.0, Fable 5.1 53.4, Qwen3.8 Max 53.4, GPT-6 Astra 52.7, GPT-6.1 Sol 51.8, Muse Spark 1.3 48.1 beside Gemini 4 Argon at 52.6: six or more labs and models at or near Google's level. Live page with no publication date (dated 'undated')."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Open-weight models handled over 70% of gateway tokens in September (DeepSeek 58%) and Moonshot took 13.2% of spend: customers route across many labs. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-september-2026","date":"2026-09-17","support":"Gemini 3 Flash lost 95% of token share since May with over three-quarters of the volume going to other labs; the same report says loyalty follows a model's profile, not its lab. Full date shown on the page."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Nearly one in two teams changed their most-used model in September. Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Claude Opus 5.5 (58), Sonnet 5.5 (56) and Meta's Muse Wonder Full (56) all score above Gemini 4 Argon (53) on Artificial Analysis. Live page with no publication date (dated 'undated')."},{"url":"https://ai.google.dev/gemini-api/docs/changelog","date":"2026-10-13","support":"The changelog calls Gemini 4 Argon, generally available 13 Oct 2026, Google's most capable model for coding, agents, math and science (company claim); it does not show a capability rivals lack. Full date shown on the page."}]}}}}},"bytedance-seed":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"bytedance-seed","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."},{"url":"https://www.khaleejtimes.com/business/tech/big-chinese-tech-firms-scramble-to-secure-huawei-ai-chips-after-deepseek-v4-launch-sources-say-2","date":"2026-04-29","support":"Reuters: Alibaba's Bailian sold DeepSeek V4 at DeepSeek's prices on release day; Tencent Cloud's TokenHub also served it (full date)."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."},{"url":"https://www.caixinglobal.com/2026-04-21/computing-shortage-forces-chinese-ai-firms-to-ration-services-102436452.html","date":"2026-04-21","support":"Caixin: compute shortage made Doubao, Alibaba Cloud, Tencent Cloud, Baidu, Zhipu, Moonshot and MiniMax ration access (full date)."},{"url":"https://eu.36kr.com/en/p/3864466789743489","date":"2026-06-23","support":"36Kr commentary: industry expects China's compute shortage to last at least two more years; cloud model prices raised (full date)."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://kr-asia.com/bytedance-raises-volcano-engines-maas-revenue-target-on-seedance-2-0-growth","date":"2026-06-05","support":"KrASIA/36Kr: Seed's coding capability 'not yet strong enough'; Moonshot, MiniMax, Tencent and telecoms sell rival coding plans (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."},{"url":"https://www.caixinglobal.com/2026-04-21/computing-shortage-forces-chinese-ai-firms-to-ration-services-102436452.html","date":"2026-04-21","support":"Caixin: compute shortage made Doubao, Alibaba Cloud, Tencent Cloud, Baidu, Zhipu, Moonshot and MiniMax ration access (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"bytedance-seed","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. Alibaba Cloud, Baidu AI Cloud and Zhipu sell competing models. Full date shown on the page."},{"url":"https://openrouter.ai/rankings","date":"undated","support":"OpenRouter's Artificial Analysis ranking (usage data to 22 Oct 2026) lists Alibaba's Qwen3.8 Max at 53.4, level with GPT-6 Astra (52.7); Alibaba says it is available through Model Studio APIs worldwide. Live page with no publication date (dated 'undated')."},{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://technode.com/2026/04/07/doubao-surpasses-120-trillion-daily-tokens-as-usage-doubles-in-three-months/","date":"2026-04-07","support":"TechNode (7 Apr 2026): Volcano Engine says Doubao exceeded 120 trillion tokens a day in March 2026, doubling in three months; with nearly half of China's MaaS token calls (IDC), rivals would have to absorb a very large volume. Full date shown on the page."},{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026): the page says chips remain the bottleneck in China, H200 purchases need case-by-case approval, and labs are moving to domestic hardware. Full date shown on the page."},{"url":"https://docs.siliconflow.cn/en/userguide/introduction","date":"undated","support":"SiliconFlow, fourth in China by token calls (IDC), says its API is compatible with the OpenAI and Anthropic chat protocols, so tools can be re-pointed at another provider by configuration. Live page with no publication date (dated 'undated')."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://www.sec.gov/Archives/edgar/data/0001577552/000110465926099220/tm2623667d1_ex99-1.htm","date":"2026-08-20","support":"Alibaba's 20 Aug 2026 filing (company statement, citing Omdia) puts Alibaba Cloud first in China's AI cloud with a 38.1% share and reports capex up 75% to RMB67.7bn in the quarter: the nearest rival is smaller than Volcano Engine in MaaS but well funded. Full date shown on the page."},{"url":"https://openrouter.ai/rankings","date":"undated","support":"Qwen3.8 Max scores 53.4 on Artificial Analysis, the best Chinese entry on the page; no Doubao Seed score was found, so capability evidence for the gap is thin. Live page with no publication date (dated 'undated')."}]}}}}},"alibaba-qwen":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"alibaba-qwen","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."},{"url":"https://technode.com/2026/04/24/deepseek-v4-preview-now-available-with-open-source-access/","date":"2026-04-24","support":"DeepSeek V4 released as open source; Pro 'approaching' some leading proprietary models; OpenAI-compatible API (full date)."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://www.alibabacloud.com/help/model-studio/deepseek-api","date":"2026-10-19","support":"Alibaba Cloud Model Studio sells DeepSeek V4 Pro and V4.1 Flash through OpenAI-compatible APIs (page updated 19 Oct 2026)."},{"url":"https://www.khaleejtimes.com/business/tech/big-chinese-tech-firms-scramble-to-secure-huawei-ai-chips-after-deepseek-v4-launch-sources-say-2","date":"2026-04-29","support":"Reuters: Alibaba's Bailian sold DeepSeek V4 at DeepSeek's prices on release day; Tencent Cloud's TokenHub also served it (full date)."},{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://www.caixinglobal.com/2026-04-21/computing-shortage-forces-chinese-ai-firms-to-ration-services-102436452.html","date":"2026-04-21","support":"Caixin: compute shortage made Doubao, Alibaba Cloud, Tencent Cloud, Baidu, Zhipu, Moonshot and MiniMax ration access (full date)."},{"url":"https://eu.36kr.com/en/p/3864466789743489","date":"2026-06-23","support":"36Kr commentary: industry expects China's compute shortage to last at least two more years; cloud model prices raised (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"alibaba-qwen","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. Volcano Engine, Baidu AI Cloud and Zhipu sell competing models. Full date shown on the page."},{"url":"https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang","date":"2026-08-21","support":"The Batch (21 Aug 2026): Kimi K3 and Z.ai's GLM-5.3 each scored 60 against Qwen3.8-Max's 58 on Artificial Analysis (the index version of that date); Alibaba released Qwen3.8-Max weights within a week, so rivals are free to host it. Full date shown on the page."},{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://www.sec.gov/Archives/edgar/data/0001577552/000110465926099220/tm2623667d1_ex99-1.htm","date":"2026-08-20","support":"Alibaba's 20 Aug 2026 filing (company statement) says AI-related product revenue rose above 100% for the twelfth straight quarter; with a 30.5% share (inputs file), ByteDance and others would need to absorb a large volume. Full date shown on the page."},{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026): the page says chips remain the bottleneck in China, H200 purchases need case-by-case approval, and labs are moving to domestic hardware. Full date shown on the page."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://openrouter.ai/rankings","date":"undated","support":"OpenRouter's Artificial Analysis ranking lists Qwen3.8 Max at 53.4 while i-scoop's best Chinese open models reach 46 on the same index: Alibaba leads those by about 7 points. No Doubao Seed score was found, so the best alternative is uncertain. Live page with no publication date (dated 'undated')."},{"url":"https://www.alibabagroup.com/document-2021044032125272064","date":"2026-08-03","support":"Alibaba says Qwen3.8-Max ranks fifth in Text Arena and fourth in Frontend Code Arena (company's own claim, 3 Aug 2026). Full date shown on the page."}]}}}}},"xai":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"xai","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"general-reasoning":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking (finance index): Claude models, GPT-6 Astra, GPT-6.1 Sol, Gemini 4 Argon, then Grok 4.7 seventh; page undated."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Launch-window spend shares: GPT-6 Astra 4.86%, Gemini 3.8 Flash 2.21%, Grok 4.7 2.00% (full date)."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Open-weight models were over 70% of September gateway tokens and nearly a quarter of spend (full date)."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Vercel says nearly one in two gateway teams changed their most-used model in September 2026, almost one in three their lab (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking (finance index): Claude models, GPT-6 Astra, GPT-6.1 Sol, Gemini 4 Argon, then Grok 4.7 seventh; page undated."},{"url":"https://www.deeplearning.ai/the-batch/sol-nearly-eclipses-astra","date":"2026-10-09","support":"Artificial Analysis Intelligence Index: Claude Opus 5.5 58, Sonnet 5.5 56, GPT-6 Astra 53, GPT-6.1 Sol 52 (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"xai","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"general-reasoning":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/leaderboards/models","date":"undated","support":"Artificial Analysis puts Grok 4.7 (xhigh) at 46, twelve points below Claude Opus 5.5 (58) and seven below GPT-6 Astra and Gemini 4 Argon (53): Anthropic, Meta, OpenAI and Google all score higher. Live page with no publication date (dated 'undated')."},{"url":"https://openrouter.ai/rankings","date":"undated","support":"OpenRouter's capability ranking (usage data to 22 Oct 2026) has ten Anthropic, Alibaba, OpenAI, Google and Meta models between 57.6 and 48.1; no xAI model appears in that top ten. Live page with no publication date (dated 'undated')."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Grok 4.7 held 2.00% of September gateway spend, and nearly one in two teams changed their most-used model that month: xAI's volume is small and customers move freely. Full date shown on the page."},{"url":"https://decrypt.co/378824/xai-launches-grok-4-7","date":"2026-09-21","support":"Decrypt, 21 Sep 2026: Grok 4.7 went live in the xAI API with no waitlist, after at least five delays since late July, priced at $2/$6 per million tokens. Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://decrypt.co/378824/xai-launches-grok-4-7","date":"2026-09-21","support":"Grok 4.7 placed second to Fable 5.1 on GDPval (1695 v 1735) and the article says Fable 5.1 leads it on CursorBench at every price point: the best alternatives exceed it. Full date shown on the page."}]}}}}},"amazon-agi":{},"mistral":{},"baidu-ernie":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"baidu-ernie","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://www.yicai.com/news/102459748.html","date":"2025-02-04","support":"Yicai: Tencent Cloud, Baidu AI Cloud, Alibaba Cloud and Volcano Engine all put DeepSeek's models on their platforms (full date)."},{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."},{"url":"https://technode.com/2026/04/24/deepseek-v4-preview-now-available-with-open-source-access/","date":"2026-04-24","support":"DeepSeek V4 released as open source; Pro 'approaching' some leading proprietary models; OpenAI-compatible API (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"baidu-ernie","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. Baidu AI Cloud is third by revenue and token calls; the two ahead and Zhipu are rivals. No ERNIE 5.x Artificial Analysis score was found, so capability evidence is thin. Full date shown on the page."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's model-management platform connects more than 300 mainstream models including DeepSeek, Qwen, Doubao and GLM; China Telecom also sells token plans: buyers choose among many models. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.siliconflow.cn/en/userguide/introduction","date":"undated","support":"SiliconFlow, fourth in China by token calls (IDC), says its API is compatible with the OpenAI and Anthropic chat protocols, so tools can be re-pointed at another provider by configuration. Live page with no publication date (dated 'undated')."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's universal token service gives access to different AI platforms under one account, its platform routes requests across 300+ models, and Tencent's workspace lets users switch models under one price: switching is a configuration step. This company's own volume is a small share of the total (inputs file; rater inference). Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."},{"url":"https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang","date":"2026-08-21","support":"The Batch (21 Aug 2026): Alibaba's Qwen3.8-Max, Moonshot's Kimi K3 and Z.ai's GLM-5.3 scored 58, 60 and 60 on Artificial Analysis (index version of that date), and Qwen3.8-Max weights are public: the leading Chinese models sit within a few points of each other. Full date shown on the page."}]}}}}},"meta-superintelligence-labs":{},"tencent-hy":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"tencent-hy","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://www.khaleejtimes.com/business/tech/big-chinese-tech-firms-scramble-to-secure-huawei-ai-chips-after-deepseek-v4-launch-sources-say-2","date":"2026-04-29","support":"Reuters: Alibaba's Bailian sold DeepSeek V4 at DeepSeek's prices on release day; Tencent Cloud's TokenHub also served it (full date)."},{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."},{"url":"https://technode.com/2026/04/24/deepseek-v4-preview-now-available-with-open-source-access/","date":"2026-04-24","support":"DeepSeek V4 released as open source; Pro 'approaching' some leading proprietary models; OpenAI-compatible API (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"tencent-hy","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. Tencent is not in IDC's top five by revenue. Full date shown on the page."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's model-management platform connects more than 300 mainstream models including DeepSeek, Qwen, Doubao and GLM; China Telecom also sells token plans: buyers choose among many models. Full date shown on the page."},{"url":"https://gigazine.net/gsc_news/en/20260707-tencent-ai-hy3/","date":"2026-07-07","support":"Gigazine (date from URL, 7 Jul 2026): Tencent released Hy3 as open weights under Apache 2.0 and says it matches open GLM-5.2 and DeepSeek-V4-Pro (company claim); its API costs 1 yuan per million input tokens. Date taken from the URL."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.siliconflow.cn/en/userguide/introduction","date":"undated","support":"SiliconFlow, fourth in China by token calls (IDC), says its API is compatible with the OpenAI and Anthropic chat protocols, so tools can be re-pointed at another provider by configuration. Live page with no publication date (dated 'undated')."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's universal token service gives access to different AI platforms under one account, its platform routes requests across 300+ models, and Tencent's workspace lets users switch models under one price: switching is a configuration step. This company's own volume is a small share of the total (inputs file; rater inference). Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."},{"url":"https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang","date":"2026-08-21","support":"The Batch (21 Aug 2026): Alibaba's Qwen3.8-Max, Moonshot's Kimi K3 and Z.ai's GLM-5.3 scored 58, 60 and 60 on Artificial Analysis (index version of that date), and Qwen3.8-Max weights are public: the leading Chinese models sit within a few points of each other. Full date shown on the page."}]}}}}},"cohere":{},"iflytek":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"iflytek","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://www.khaleejtimes.com/business/tech/big-chinese-tech-firms-scramble-to-secure-huawei-ai-chips-after-deepseek-v4-launch-sources-say-2","date":"2026-04-29","support":"Reuters: DeepSeek V4 runs on Huawei Ascend chips, lifting demand; Ascend 950 output expected to fall short of demand (full date)."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."},{"url":"https://www.alibabacloud.com/help/model-studio/deepseek-api","date":"2026-10-19","support":"Alibaba Cloud Model Studio sells DeepSeek V4 Pro and V4.1 Flash through OpenAI-compatible APIs (page updated 19 Oct 2026)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."},{"url":"https://www.khaleejtimes.com/business/tech/big-chinese-tech-firms-scramble-to-secure-huawei-ai-chips-after-deepseek-v4-launch-sources-say-2","date":"2026-04-29","support":"Reuters: DeepSeek V4 runs on Huawei Ascend chips, lifting demand; Ascend 950 output expected to fall short of demand (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"iflytek","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. iFlytek is not in IDC's top five by revenue or token calls; no iFlytek-specific capability source was found, so this rests on market structure. Full date shown on the page."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's model-management platform connects more than 300 mainstream models including DeepSeek, Qwen, Doubao and GLM; China Telecom also sells token plans: buyers choose among many models. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.siliconflow.cn/en/userguide/introduction","date":"undated","support":"SiliconFlow, fourth in China by token calls (IDC), says its API is compatible with the OpenAI and Anthropic chat protocols, so tools can be re-pointed at another provider by configuration. Live page with no publication date (dated 'undated')."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's universal token service gives access to different AI platforms under one account, its platform routes requests across 300+ models, and Tencent's workspace lets users switch models under one price: switching is a configuration step. This company's own volume is a small share of the total (inputs file; rater inference). Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."},{"url":"https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang","date":"2026-08-21","support":"The Batch (21 Aug 2026): Alibaba's Qwen3.8-Max, Moonshot's Kimi K3 and Z.ai's GLM-5.3 scored 58, 60 and 60 on Artificial Analysis (index version of that date), and Qwen3.8-Max weights are public: the leading Chinese models sit within a few points of each other. Full date shown on the page."}]}}}}},"typesafe-ai":{},"sber":{},"perplexity":{},"naver-hyperclova":{},"moonshot":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"moonshot","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://huggingface.co/moonshotai/Kimi-K3","date":"undated","support":"Moonshot publishes Kimi K3 as 'an open-weight, native multimodal agentic model' under the Kimi K3 License; page undated."},{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://huggingface.co/moonshotai/Kimi-K3","date":"undated","support":"Moonshot publishes Kimi K3 as 'an open-weight, native multimodal agentic model' under the Kimi K3 License; page undated."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"moonshot","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. Moonshot is not in IDC's top five; Kimi K3 is open-weight. Full date shown on the page."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's model-management platform connects more than 300 mainstream models including DeepSeek, Qwen, Doubao and GLM; China Telecom also sells token plans: buyers choose among many models. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.siliconflow.cn/en/userguide/introduction","date":"undated","support":"SiliconFlow, fourth in China by token calls (IDC), says its API is compatible with the OpenAI and Anthropic chat protocols, so tools can be re-pointed at another provider by configuration. Live page with no publication date (dated 'undated')."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Vercel (15 Oct 2026): Moonshot took 13.2% of September gateway spend while nearly one in two teams changed their most-used model that month: Moonshot gained share in weeks and could lose it as quickly. Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."},{"url":"https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang","date":"2026-08-21","support":"The Batch (21 Aug 2026): Alibaba's Qwen3.8-Max, Moonshot's Kimi K3 and Z.ai's GLM-5.3 scored 58, 60 and 60 on Artificial Analysis (index version of that date), and Qwen3.8-Max weights are public: the leading Chinese models sit within a few points of each other. Full date shown on the page."}]}}}}},"sensetime":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"sensetime","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."},{"url":"https://www.khaleejtimes.com/business/tech/big-chinese-tech-firms-scramble-to-secure-huawei-ai-chips-after-deepseek-v4-launch-sources-say-2","date":"2026-04-29","support":"Reuters: Alibaba's Bailian sold DeepSeek V4 at DeepSeek's prices on release day; Tencent Cloud's TokenHub also served it (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."},{"url":"https://technode.com/2026/04/24/deepseek-v4-preview-now-available-with-open-source-access/","date":"2026-04-24","support":"DeepSeek V4 released as open source; Pro 'approaching' some leading proprietary models; OpenAI-compatible API (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"sensetime","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. IDC names SenseTime among top private-deployment vendors, not among the top five public-cloud vendors. Full date shown on the page."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's model-management platform connects more than 300 mainstream models including DeepSeek, Qwen, Doubao and GLM; China Telecom also sells token plans: buyers choose among many models. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.siliconflow.cn/en/userguide/introduction","date":"undated","support":"SiliconFlow, fourth in China by token calls (IDC), says its API is compatible with the OpenAI and Anthropic chat protocols, so tools can be re-pointed at another provider by configuration. Live page with no publication date (dated 'undated')."},{"url":"https://technode.global/2026/08/27/sensetime-h1-2026-first-profit-generative-ai-revenue/","date":"2026-08-27","support":"TechNode (27 Aug 2026): SenseTime's platform averaged 2.4 trillion tokens a day in July 2026 and serves four external foundation-model providers; the article does not split API from private-deployment revenue, so switching evidence for the API line is thin. Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."},{"url":"https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang","date":"2026-08-21","support":"The Batch (21 Aug 2026): Alibaba's Qwen3.8-Max, Moonshot's Kimi K3 and Z.ai's GLM-5.3 scored 58, 60 and 60 on Artificial Analysis (index version of that date), and Qwen3.8-Max weights are public: the leading Chinese models sit within a few points of each other. Full date shown on the page."}]}}}}},"microsoft-ai":{},"china-mobile":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"china-mobile","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://www.lightreading.com/ai-machine-learning/china-telcos-rush-to-embrace-deepseek","date":"2025-02-11","support":"China Mobile's cloud offers full DeepSeek V1-V3 and R1; its own Jiutian models target industry verticals (full date)."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://www.lightreading.com/ai-machine-learning/china-telcos-rush-to-embrace-deepseek","date":"2025-02-11","support":"China Mobile's cloud offers full DeepSeek V1-V3 and R1; its own Jiutian models target industry verticals (full date)."},{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.lightreading.com/ai-machine-learning/china-telcos-rush-to-embrace-deepseek","date":"2025-02-11","support":"China Mobile's cloud offers full DeepSeek V1-V3 and R1; its own Jiutian models target industry verticals (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"china-mobile","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. China Mobile Cloud is fifth by revenue and by token calls, the smallest of IDC's five; no Jiutian capability score was found. Full date shown on the page."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's model-management platform connects more than 300 mainstream models including DeepSeek, Qwen, Doubao and GLM; China Telecom also sells token plans: buyers choose among many models. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.siliconflow.cn/en/userguide/introduction","date":"undated","support":"SiliconFlow, fourth in China by token calls (IDC), says its API is compatible with the OpenAI and Anthropic chat protocols, so tools can be re-pointed at another provider by configuration. Live page with no publication date (dated 'undated')."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's universal token service gives access to different AI platforms under one account, its platform routes requests across 300+ models, and Tencent's workspace lets users switch models under one price: switching is a configuration step. This company's own volume is a small share of the total (inputs file; rater inference). Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."},{"url":"https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang","date":"2026-08-21","support":"The Batch (21 Aug 2026): Alibaba's Qwen3.8-Max, Moonshot's Kimi K3 and Z.ai's GLM-5.3 scored 58, 60 and 60 on Artificial Analysis (index version of that date), and Qwen3.8-Max weights are public: the leading Chinese models sit within a few points of each other. Full date shown on the page."}]}}}}},"zhipu":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"zhipu","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://kr-asia.com/bytedance-raises-volcano-engines-maas-revenue-target-on-seedance-2-0-growth","date":"2026-06-05","support":"KrASIA/36Kr: Seed's coding capability 'not yet strong enough'; Moonshot, MiniMax, Tencent and telecoms sell rival coding plans (full date)."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://simonwillison.net/2026/Jun/17/glm-52/","date":"2026-06-17","support":"Z.ai released full GLM-5.2 weights under an MIT licence; second on Code Arena WebDev behind Claude Fable 5 (full date)."},{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://simonwillison.net/2026/Jun/17/glm-52/","date":"2026-06-17","support":"Z.ai released full GLM-5.2 weights under an MIT licence; second on Code Arena WebDev behind Claude Fable 5 (full date)."},{"url":"https://artificialanalysis.ai/","date":"undated","support":"Capability ranking lists Kimi K3, Qwen3.8 Max, Step 5 Preview and GLM-5.3 near each other; no Doubao Seed model in top 15; undated."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"zhipu","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. Zhipu is fourth by revenue; its flagship GLM models are open-weight. Full date shown on the page."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's model-management platform connects more than 300 mainstream models including DeepSeek, Qwen, Doubao and GLM; China Telecom also sells token plans: buyers choose among many models. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.siliconflow.cn/en/userguide/introduction","date":"undated","support":"SiliconFlow, fourth in China by token calls (IDC), says its API is compatible with the OpenAI and Anthropic chat protocols, so tools can be re-pointed at another provider by configuration. Live page with no publication date (dated 'undated')."},{"url":"https://www.caixinglobal.com/2026-09-01/zhipu-revenue-rises-nearly-400-as-cloud-business-gains-ground-102480471.html","date":"2026-09-01","support":"Caixin (1 Sep 2026): Zhipu's shares fell 28.5% on the day rival Moonshot launched Kimi K3, and Zhipu answered with GLM-5.3 in August: buyers and investors treat the labs as close substitutes. Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."},{"url":"https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang","date":"2026-08-21","support":"The Batch (21 Aug 2026): Alibaba's Qwen3.8-Max, Moonshot's Kimi K3 and Z.ai's GLM-5.3 scored 58, 60 and 60 on Artificial Analysis (index version of that date), and Qwen3.8-Max weights are public: the leading Chinese models sit within a few points of each other. Full date shown on the page."}]}}}}},"huawei-cloud":{"a":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"huawei-cloud","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: Volcano Engine over 40% of 2025 MaaS revenue, then Alibaba, Baidu, Zhipu, China Mobile; Tencent, SenseTime, Huawei also named (full date)."},{"url":"https://www.khaleejtimes.com/business/tech/big-chinese-tech-firms-scramble-to-secure-huawei-ai-chips-after-deepseek-v4-launch-sources-say-2","date":"2026-04-29","support":"Reuters: DeepSeek V4 runs on Huawei Ascend chips, lifting demand; Ascend 950 output expected to fall short of demand (full date)."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC: buyers mainly chose vendors on price in the past two years and often call model APIs directly without fine-tuning (full date)."},{"url":"https://www.yicai.com/news/102459748.html","date":"2025-02-04","support":"Yicai: Tencent Cloud, Baidu AI Cloud, Alibaba Cloud and Volcano Engine all put DeepSeek's models on their platforms (full date)."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.khaleejtimes.com/business/tech/big-chinese-tech-firms-scramble-to-secure-huawei-ai-chips-after-deepseek-v4-launch-sources-say-2","date":"2026-04-29","support":"Reuters: DeepSeek V4 runs on Huawei Ascend chips, lifting demand; Ascend 950 output expected to fall short of demand (full date)."},{"url":"https://technode.com/2026/04/24/deepseek-v4-preview-now-available-with-open-source-access/","date":"2026-04-24","support":"DeepSeek V4 released as open source; Pro 'approaching' some leading proprietary models; OpenAI-compatible API (full date)."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-frontier-labs","entityId":"huawei-cloud","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"china-api":{"alternatives":{"level":1,"evidence":[{"url":"https://www.idc.com/resource-center/blog/%E4%B8%AD%E5%9B%BDmaas%E5%B8%82%E5%9C%BA%E8%BF%9B%E5%85%A5%E9%AB%98%E9%80%9F%E5%A2%9E%E9%95%BF%E6%9C%9F%EF%BC%8Ctoken%E7%BB%8F%E6%B5%8E%E4%BB%8E%E6%A6%82%E5%BF%B5%E8%B5%B0%E5%90%91%E8%A7%84%E6%A8%A1/","date":"2026-05-07","support":"IDC China MaaS blog (7 May 2026), 2025: Volcano Engine first by revenue (over 40%) and by token calls (nearly half); Alibaba Cloud second; Baidu AI Cloud third; Zhipu fourth by revenue; China Mobile Cloud fifth on both lists and SiliconFlow fourth by token calls. Huawei Cloud is not among IDC's top five by revenue or token calls; no 2026 Pangu capability or usage source was found, so this rests on market structure. Full date shown on the page."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's model-management platform connects more than 300 mainstream models including DeepSeek, Qwen, Doubao and GLM; China Telecom also sells token plans: buyers choose among many models. Full date shown on the page."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.siliconflow.cn/en/userguide/introduction","date":"undated","support":"SiliconFlow, fourth in China by token calls (IDC), says its API is compatible with the OpenAI and Anthropic chat protocols, so tools can be re-pointed at another provider by configuration. Live page with no publication date (dated 'undated')."},{"url":"https://govt.chinadaily.com.cn/s/202605/20/WS6a0d59e4498e36855033cb18/telecom-giants-launch-ai-token-services.html","date":"2026-05-20","support":"China Daily (20 May 2026): China Mobile's universal token service gives access to different AI platforms under one account, its platform routes requests across 300+ models, and Tencent's workspace lets users switch models under one price: switching is a configuration step. This company's own volume is a small share of the total (inputs file; rater inference). Full date shown on the page."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.i-scoop.eu/chinese-ai-labs/","date":"2026-10-18","support":"i-scoop (18 Oct 2026), Artificial Analysis scores: Xiaomi MiMo-V2.6-Pro 46, Z.ai GLM-5.3 45, StepFun Step 5 Preview 44, Moonshot Kimi K3 44 (all open-weight), against 58 for Claude Opus 5.5. Full date shown on the page."},{"url":"https://www.deeplearning.ai/the-batch/qwen3-8-max-lands-with-a-bang","date":"2026-08-21","support":"The Batch (21 Aug 2026): Alibaba's Qwen3.8-Max, Moonshot's Kimi K3 and Z.ai's GLM-5.3 scored 58, 60 and 60 on Artificial Analysis (index version of that date), and Qwen3.8-Max weights are public: the leading Chinese models sit within a few points of each other. Full date shown on the page."}]}}}}},"stepfun":{},"kuaishou-kwaipilot":{},"writer":{},"ntt":{},"upstage":{},"minimax":{},"inception-labs":{}},"validation":[]}