AI Lab Tracker — what each AI company shipped lately (updated Oct 21, 2026)

Releases from roughly the last 30 days (2026-09-21 to 2026-10-21). Every item links to its official source. Benchmark numbers marked vendor-reported come from the company itself. Third-party figures come from public leaderboards and show their snapshot date. We have not run hands-on tests. Machine-readable version: `ai-labs.json`.

This week's top 5

1. Google's Gemini 4 Argon (Oct 13) debuted at #1 on LMArena Text (1524, marked preliminary) and scores 53 on the Artificial Analysis Intelligence Index. 2. Anthropic's Claude 5.5 family: Opus 5.5 (Sep 22) leads Artificial Analysis (58, max with fallback); Sonnet 5.5 (Sep 28) is #2 at 56; Haiku 5.5 (Oct 7) launched at $0.10/$0.50 per M tokens for prompts up to 100k. 3. Meta's Muse wave: Muse Wonder / Wonder Full (Oct 12; Wonder is #2 on LMArena, Wonder Full scores 56 on AA), open-weight Muse Spark Open (Oct 15) and Apache-2.0 Muse Glimmer 30B (Oct 19). 4. OpenAI shipped GPT-6.1 Sol (Sep 29; AA 52 at max, $2/$10) and rolled GPT-6 out to all ChatGPT users with Intelligent UI (Oct 7). 5. Cheap Chinese agent models got stronger: DeepSeek-V4.1-Flash-1017 (Oct 17, MIT, AA 45, the #2 open-weights model) and MiniMax M3.1 Flash (Oct 19, $0.15/$0.60, AA 43).

Leaderboard snapshots: LMArena Text Arena (overall) (2026-10-20); Artificial Analysis Intelligence Index (2026-10-21)

OpenAI

### GPT-6 + Intelligent UI rolled out to everyone in ChatGPT - Date: 2026-10-07 · Official link: https://openai.com/index/gpt-6-for-everyone/ - What shipped: GPT-6 is rolling out globally in ChatGPT on free and paid plans. Intelligent UI lets answers include graphics, buttons, forms and charts. - Checked 2026-10-21

### GPT-6.1 Sol - Date: 2026-09-29 · Official link: https://openai.com/index/introducing-gpt-6-1-sol/ - What shipped: An upgrade to GPT-6 Sol. OpenAI says it nearly matches GPT-6 Astra on agentic coding, computer use and professional work. API id gpt-6.1-sol, priced at $2 input / $0.10 cached / $10 output per M tokens. Available in ChatGPT Work and Codex for paid plans. - How it performs: - vendor-reported: Matches GPT-6 Astra on DeepSWE v1.1 at about one-fifth of the cost; 6.4 pp above GPT-6 Sol's best score. Source: official page, 2026-09-29 - vendor-reported: AutomationBench: 2.2 pp above Claude Opus 5.5 at medium effort. Source: official page, 2026-09-29 - third-party: Artificial Analysis Intelligence Index: 52 (max), 51 (xhigh); cost per task $0.72 at max. Source: Artificial Analysis Intelligence Index, 2026-10-21 - third-party: LMArena Text: #28, score 1479±8 (Max). Source: LMArena Text Arena (overall), 2026-10-20 - Checked 2026-10-21

### GPT-6 Sol and GPT-6 Luna - Date: 2026-09-22 · Official link: https://openai.com/index/introducing-gpt-6-sol-and-luna/ - What shipped: Cheaper models from the GPT-6 family. API prices: Sol $2/$10, Luna $0.10/$0.50 per M tokens, 50% below GPT-5.6 Sol/Luna promo prices. - How it performs: - vendor-reported: AutomationBench 1.0.6: GPT-6 Sol (xhigh) 33.2% vs Claude Opus 5 (max) 26.9%. Source: official page, 2026-09-22 - third-party: LMArena Text: GPT-6 Sol (Max) #62 (1459±6); GPT-6 Luna (Max) #96 (1441±4). Source: LMArena Text Arena (overall), 2026-10-20 - third-party: Artificial Analysis: GPT-6 Luna (max) 38. Source: Artificial Analysis Intelligence Index, 2026-10-21 - Note: OpenAI's release index dates this to Sep 22, 2026. A search snippet showed a 2026-10-15 'published' timestamp on the page, so it may have been republished. We use Sep 22. - Checked 2026-10-21

Anthropic

### Claude Haiku 5.5 - Date: 2026-10-07 · Official link: https://www.anthropic.com/claude-haiku-5-5 - What shipped: A small, fast model (claude-haiku-5-5) with a 1M-token context, 128k max output and an adjustable effort setting. Prices per M tokens: $0.10 input / $0.50 output for prompts up to 100k, and $0.50 / $2.50 above that. The same day, Anthropic halved Sonnet 5.5 cache-read prices and added monthly API credits to Max and Team plans. - How it performs: - vendor-reported: Terminal-Bench 4.0: 39.2% (Haiku 4.5: 0.0%, GPT-6 Luna: 16.4%). Source: official page, 2026-10-07 - vendor-reported: OSWorld 2.1 (offline subset): 72.4%. Source: official page, 2026-10-07 - third-party: Artificial Analysis Intelligence Index: 43 (max), 41 (xhigh). Source: Artificial Analysis Intelligence Index, 2026-10-21 - Checked 2026-10-21

### Claude Code Review generally available (listing only) - Date: 2026-10-15 · Official link: https://claude.com/resources/product-announcements - What shipped: Code Review moved to general availability. - Note: Confirmed only from the announcements listing; we did not open the full post. - Checked 2026-10-21

### Claude Dashboards and Claude Motion (beta); Docs, Slides and Design GA (listing only) - Date: 2026-10-08 · Official link: https://claude.com/resources/product-announcements - What shipped: Dashboards and Motion entered beta. Docs, Slides and Design left beta and are on every plan, including Free. - Checked 2026-10-21

### Claude Sonnet 5.5 - Date: 2026-09-28 · Official link: https://www.anthropic.com/news/claude-sonnet-5-5 - What shipped: The second model in the Claude 5.5 family (claude-sonnet-5-5). Priced at $2/$10 per M tokens. Anthropic says it is 30%+ faster than Sonnet 5 and up to 30% cheaper per task. - How it performs: - vendor-reported: Terminal-Bench 4.0: 70.6% (Sonnet 5: 10.3%). Source: official page, 2026-09-28 - vendor-reported: CursorBench 4.0: 55.5%; OSWorld 2.1 partial: 80.1%. Source: official page, 2026-09-28 - third-party: Artificial Analysis Intelligence Index: 56 (max with fallback), #2 overall. Source: Artificial Analysis Intelligence Index, 2026-10-21 - third-party: LMArena Text: #30, 1477±7 (xHigh). Source: LMArena Text Arena (overall), 2026-10-20 - Checked 2026-10-21

### Claude Opus 5.5 - Date: 2026-09-22 · Official link: https://www.anthropic.com/news - What shipped: Anthropic says Opus 5.5 leads in agentic coding and knowledge work and costs 40% less to run than Opus 5 on typical workloads. It is priced at $4/$20 per M tokens (from the Sonnet 5.5 pricing table). - How it performs: - vendor-reported: Terminal-Bench 4.0: 66.4% (xhigh), as reported in the Sonnet 5.5 post. Source: official page, 2026-09-28 - third-party: Artificial Analysis Intelligence Index: 58 (max with fallback), #1 overall. Source: Artificial Analysis Intelligence Index, 2026-10-21 - third-party: LMArena Text: #3, 1509±7 (High). Source: LMArena Text Arena (overall), 2026-10-20 - Note: Date and summary come from the Anthropic newsroom listing; we did not open the Opus 5.5 launch post itself. - Checked 2026-10-21

Google / DeepMind

### Gemini 4 Argon - Date: 2026-10-13 · Official link: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ - What shipped: Google's new frontier model (gemini-4-argon), built for coding, agentic work, math, science, finance, legal and cybersecurity tasks. It is available through the Gemini API and AI Studio, in Antigravity for Ultra subscribers, and in the Gemini app for Pro and Ultra. Introductory price is $2/$10 per M tokens through Nov 30, 2026, then $4/$20. - How it performs: - vendor-reported: Ranked #1 on the Vals Index (per Google). Source: official page, 2026-10-13 - third-party: LMArena Text: #1, 1524±9 (High), marked Preliminary. Source: LMArena Text Arena (overall), 2026-10-20 - third-party: Artificial Analysis Intelligence Index: 53 (high). Source: Artificial Analysis Intelligence Index, 2026-10-21 - Checked 2026-10-21

xAI (SpaceXSI)

### Grok Voice 3.0 - Date: 2026-10-14 · Official link: https://x.ai/news/grok-voice-3 - What shipped: A speech-to-speech voice model (grok-voice-3.0). The grok-voice-latest alias now points to it. Priced at $0.08/min, the same as Grok Voice 2.0. - How it performs: - third-party (as quoted by vendor): #1 on the Artificial Analysis Speech-to-Speech Index at 86.9 (AA, Oct 13, 2026, as quoted by xAI). Source: official page, 2026-10-14 - Note: We have not yet checked the AA figure on Artificial Analysis's own speech leaderboard. - Checked 2026-10-21

### The X Subscription (Grok + Grok Bot + Cursor + X Premium) - Date: 2026-10-12 · Official link: https://x.ai/news/x-subscription - What shipped: One plan that bundles Grok, Grok Bot, Cursor and X Premium, with a shared pool of Grok model usage across Grok and Cursor. - Checked 2026-10-21

### Grok 4.7 - Date: 2026-09-22 · Official link: https://x.ai/news/grok-4-7 - What shipped: xAI's coding and knowledge-work model, built on a new, larger base model. Priced at $2/$6 per M tokens, with a fast variant at twice the price. Available in Cursor, Grok Build and the API. - How it performs: - vendor-reported: CursorBench 4.0: 46.3%; Terminal-Bench 4.0: 37.6%; DeepSWE v1.1: 71.0% (high effort). Source: official page, 2026-09-22 - third-party: Artificial Analysis Intelligence Index: 46 (xhigh/high). Source: Artificial Analysis Intelligence Index, 2026-10-21 - third-party: LMArena Text: #91, 1442±6 (xHigh). Source: LMArena Text Arena (overall), 2026-10-20 - Note: The Sep 22 date comes from the x.ai/news listing. - Checked 2026-10-21

Mistral

### Voxtral 3 (speech-to-text, text-to-speech, voice agent SDK) (date Unverified) - Date: 2026-10-20 · Official link: https://mistral.ai/news/voxtral-3/ - What shipped: New transcription models covering 50+ languages, Voxtral Mini TTS 3 (35+ languages, $13 per M characters), and a voice agent SDK built on LiveKit. Available through the EU-hosted API or self-deployment. - How it performs: - vendor-reported: Mistral says Voxtral Small Transcribe 3 outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe on multilingual transcription. Source: official page, 2026-10-20 - Note: The page shows no publish date. Oct 20 comes from search results and chart footers marked 'generated 2026-10-20'. - Checked 2026-10-21

### Timestral (time-series foundation models) - Date: 2026-10-15 · Official link: https://mistral.ai/news/timestral/ - What shipped: Mistral's first time-series forecasting models, in 5M, 28M, 123M and 209M sizes, under a Modified MIT license. They run on CPU. - How it performs: - vendor-reported: 209M model ranks first on GIFT-Eval, TIME and fev-bench among zero-shot pretrained models (results submitted to the leaderboards). Source: official page, 2026-10-15 - Checked 2026-10-21

### Mistral Large 4 (public preview) - Date: 2026-10-06 · Official link: https://mistral.ai/news/mistral-large-4/ - What shipped: A 1T-parameter (52B active) multimodal MoE in public preview. Mistral promised open weights by the end of October. - How it performs: - vendor-reported: DeepSWE v1.1 61.7%, Terminal-Bench 4 28.3%, AutomationBench 59.9%. Source: official page, 2026-10-06 - third-party: Artificial Analysis Intelligence Index: 38 (Preview). Source: Artificial Analysis Intelligence Index, 2026-10-21 - third-party: LMArena Text: #112, 1433±7. Source: LMArena Text Arena (overall), 2026-10-20 - Checked 2026-10-21

DeepSeek

### DeepSeek-V4.1-Flash-1017 + Search/Fetch API opened - Date: 2026-10-17 · Official link: https://api-docs.deepseek.com/updates/ - What shipped: A minor update to V4.1-Flash, served through the API as deepseek-flash at unchanged prices. MIT-licensed weights are on Hugging Face. DeepSeek also opened its in-house search API (Search & Fetch). - How it performs: - vendor-reported: Terminal-Bench 4.0: 42.3 (prev 31.2); DeepSWE v1.1: 77.0; HLE w/o tools: 42.6. Source: official page, 2026-10-17 - third-party: Artificial Analysis Intelligence Index: 45 (max), #2 open-weights model. Source: Artificial Analysis Intelligence Index, 2026-10-21 - Note: We did not confirm the Search/Fetch API price ($0.75 per 1,000 calls) on an official page, so it is left out. - Checked 2026-10-21

Alibaba Qwen

### Qwen-Image-2.1-Turbo (open source) - Date: 2026-10-09 · Official link: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/newly-released-models - What shipped: The open-source, cost-efficient model in the Qwen-Image 2.1 series. It combines generation and editing, supports transparent output and accepts up to 10 reference images. - Checked 2026-10-21

### Qwen-Image-2.1-Pro - Date: 2026-10-02 · Official link: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/newly-released-models - What shipped: Text-to-image and editing in one model, with a 7B visual generation component and native transparency. - Note: No new Qwen LLM appeared in the window. The latest is Qwen3.8-Max-0902 (Sep 2), which LMArena ranks #23 (1483±5). - Checked 2026-10-21

MiniMax

### MiniMax M3.1 Flash - Date: 2026-10-19 · Official link: https://www.minimax.io/blog/minimax-m31f - What shipped: A fast, low-cost agent and coding model priced at $0.15/$0.60 per M tokens ($0.015 cached). MiniMax says weights will be released later. - How it performs: - vendor-reported: DeepSWE v1.1 73.7%; OSWorld V2 61.7; Agents' Last Exam 29.6; CyberGym 82%. Source: official page, 2026-10-19 - third-party: Artificial Analysis Intelligence Index: 43. Source: Artificial Analysis Intelligence Index, 2026-10-21 - Note: A third-party summary said weights would come on Oct 23. The official post only says 'later', so that date is Unverified. - Checked 2026-10-21

Moonshot / Kimi

### No official release in the window - Date: — · Official link: https://lmarena.ai/leaderboard/text - What shipped: The latest official flagship is still Kimi K3 (LMArena #18, 1488±5 at Max; AA 44). Reports of a 'Kimi K3.1' preview are leaks or rumors with no official page. - Note: Kimi K3.1 is Unverified and not listed as a release. - Checked 2026-10-21

Others notable

No other lab release in the window was verified against an official page this cycle. Candidates to check next time: Z.ai GLM, Xiaomi MiMo, StepFun, Tencent Hy.

Unverified or dropped this cycle

Mistral 'Ministral 3B/8B on Oct 9': a search summary misread the 2024 changelog entry
Kimi K3.1: leaks/rumors only
MiniMax M3.1 Flash weights 'Oct 23': not stated officially
DeepSeek Search API price: no official page fetched

--- Methodology: we read each official announcement and quote figures from it, plus public leaderboard snapshots. We do not run our own tests. Figures change often, so check the linked sources.