The scoreboard of the three paid US labs and the seven Chinese labs that weigh most in traffic, and what each one logged in its own release notes: new models, pricing, deprecations and API features. The numbers come from real token traffic; the moves, from the lab’s own official page.
In the week of Oct 12, 2026, the three paid US labs took 19.5% of tokens and 63.0% of estimated spend. Chinese labs took 68.9% of tokens and 35.8% of spend.
The three paid US labs
Tokens
19.5%
↓ 1.1 pp
Estimated spend
63.0%
↓ 8.3 pp
tokensestimated spend13 weeks; arrow vs. 4 weeks earlier
Anthropic, OpenAI and Google
Chinese labs
Tokens
68.9%
↑ 0.1 pp
Estimated spend
35.8%
↑ 8.3 pp
tokensestimated spend13 weeks; arrow vs. 4 weeks earlier
Every lab of Chinese origin in the data. The seven tracked here add up to 60.8% of tokens.
The three paid US labs
Anthropic
#4 by tokens · week of Oct 12, 2026
Tokens
7.9%
↑ 4.6 pp
Estimated spend
45.2%
↑ 20.8 pp
tokensestimated spend13 weeks; arrow vs. 4 weeks earlier
Fast mode for Claude Opus 5.5 is now open to everyone with no access request, and it gets a default rate limit.
Technical version
Requests to Claude Opus 5.5 no longer need the fast-mode-2026-02-01 beta header, and requests that still send it keep working. Fast mode for Claude Opus 5 and Claude Opus 4.8 stays in research preview.
Anthropic launched Claude Haiku 5.5, which it calls its most capable model for high-volume, latency-sensitive work.
Technical version
claude-haiku-5-5 has a 1M token context window, up to 128k output tokens and adaptive thinking set by the effort parameter. Available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Reusing cached context on Claude Sonnet 5.5 got cheaper, and every other price stayed the same.
Technical version
Prompt cache reads drop from $0.20 to $0.10 per million tokens, 0.05x the base input price instead of 0.1x. Cache writes and other prices are unchanged.
OpenAI's Decisions API is now generally available, with shared input caching and more questions per request.
Technical version
Available with gpt-6-luna. Each request takes up to 400 questions, up from 200 in public beta, and accepts structured data directly through input_json.
GPT-6.1 Sol users can ask for faster token generation, with its own pricing.
Technical version
Use gpt-6.1-sol with service_tier: "ultrafast" to cut the time between output tokens. Open to all API users, with global processing and US and EU data residency.
OpenAI released GPT-6.1 Sol for complex coding and professional work, costing less than GPT-6 Astra.
Technical version
gpt-6.1-sol: $2 input, $0.10 cached input, $2.50 cache write and $10 output per 1M tokens for prompts up to 272K tokens. Multi-agent is in beta in the Responses API.
The five Chinese labs with the most traffic over 90 days, plus Moonshot (Kimi) and Qwen (Alibaba). The list is reviewed every quarter from the data, not from preference.
DeepSeek
#1 by tokens · week of Oct 12, 2026
Tokens
30.1%
↑ 3.1 pp
Estimated spend
7.4%
↑ 2.4 pp
tokensestimated spend13 weeks; arrow vs. 4 weeks earlier
DeepSeek released a new V4.1 Flash version that it says scores higher, finishes tasks faster and writes shorter answers.
Technical version
Live on the website, app and API, optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. DeepSeek recommends DeepSeek Harness v0.2.1-rc.1 or later.
Z.ai released GLM-5.3 and says it codes much better than GLM-5.2 and finds security flaws in code.
Technical version
Z.ai reports a 50% gain over GLM-5.2 on Z.ai Code Bench and state of the art among open models on public benchmarks such as Terminal Bench 3.0. On real targets it found 2,436 vulnerabilities, 1,097 of medium or high severity.
Alibaba now serves qwen3.8-max and qwen3.7-flash from the US (Virginia) region.
Technical version
US deployment scope with text generation, deep thinking, visual understanding, function calling, structured output and context caching. qwen3.8-max has a native 1M token context window.
OpenAI's Decisions API is now generally available, with shared input caching and more questions per request.
Technical version
Available with gpt-6-luna. Each request takes up to 400 questions, up from 200 in public beta, and accepts structured data directly through input_json.
DeepSeek released a new V4.1 Flash version that it says scores higher, finishes tasks faster and writes shorter answers.
Technical version
Live on the website, app and API, optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. DeepSeek recommends DeepSeek Harness v0.2.1-rc.1 or later.
Fast mode for Claude Opus 5.5 is now open to everyone with no access request, and it gets a default rate limit.
Technical version
Requests to Claude Opus 5.5 no longer need the fast-mode-2026-02-01 beta header, and requests that still send it keep working. Fast mode for Claude Opus 5 and Claude Opus 4.8 stays in research preview.
Alibaba now serves qwen3.8-max and qwen3.7-flash from the US (Virginia) region.
Technical version
US deployment scope with text generation, deep thinking, visual understanding, function calling, structured output and context caching. qwen3.8-max has a native 1M token context window.
GPT-6.1 Sol users can ask for faster token generation, with its own pricing.
Technical version
Use gpt-6.1-sol with service_tier: "ultrafast" to cut the time between output tokens. Open to all API users, with global processing and US and EU data residency.
Anthropic launched Claude Haiku 5.5, which it calls its most capable model for high-volume, latency-sensitive work.
Technical version
claude-haiku-5-5 has a 1M token context window, up to 128k output tokens and adaptive thinking set by the effort parameter. Available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Reusing cached context on Claude Sonnet 5.5 got cheaper, and every other price stayed the same.
Technical version
Prompt cache reads drop from $0.20 to $0.10 per million tokens, 0.05x the base input price instead of 0.1x. Cache writes and other prices are unchanged.
OpenAI released GPT-6.1 Sol for complex coding and professional work, costing less than GPT-6 Astra.
Technical version
gpt-6.1-sol: $2 input, $0.10 cached input, $2.50 cache write and $10 output per 1M tokens for prompts up to 272K tokens. Multi-agent is in beta in the Responses API.
Anthropic launched Claude Sonnet 5.5, available on its API and on the Amazon, Google and Microsoft platforms.
Technical version
claude-sonnet-5-5 is on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. Context window, output limits and prices are on the model page.
Anthropic launched Claude Opus 5.5, a model for long-running coding and knowledge work, priced below Claude Opus 5.
Technical version
claude-opus-5-5 has a 1M token context window, up to 128k output tokens and always-on adaptive thinking, at $4 / $20 per MTok versus $5 / $25 for Claude Opus 5.
Z.ai released GLM-5.3 and says it codes much better than GLM-5.2 and finds security flaws in code.
Technical version
Z.ai reports a 50% gain over GLM-5.2 on Z.ai Code Bench and state of the art among open models on public benchmarks such as Terminal Bench 3.0. On real targets it found 2,436 vulnerabilities, 1,097 of medium or high severity.
Moonshot put Kimi K3, its flagship for long-horizon coding and end-to-end knowledge work, on its API.
Technical version
1M token context. New users no longer get kimi-k2.5 or the moonshot-v1 series.
Token and estimated spend share among identified models, per full week, Monday to Sunday, the same math as the “The data” strip in the Radar. Spend is tokens times the catalog blended price (75% input, 25% output), and free endpoints count as zero. Moves are AI-written summaries using only what the official note says, with every number checked against it; collection runs Monday, Wednesday and Friday. ChatGPT release notes refuse automated reading, so OpenAI is covered through its API changelog.