Frontier AI model releases, by lab
A dated record of the frontier AI models ten labs have made public since ChatGPT: anyone could use them on the date given, public previews and betas included. Partner-only and announced-only models are left out. A lab’s top two tiers count, as does any tier sold above them; the smallest of three or more never does. Which plans and apps each new model reached, and when, is in The Access Gap.
The strip and the list are the record. The chart at the foot of the page is the argument this publication makes from it, drawn by hand and labelled that way.
01 · The strip
When each lab shipped
Ten labs shipped 75 frontier releases in the year to October 2026, up from 53
- OpenAI 23 GPT-6.1 Terra26 Oct 2026
- Anthropic 21 Claude Sonnet 5.528 Sep 2026
- Google 19 Gemini 4 Argon (GA)13 Oct 2026
- DeepSeek 19 DeepSeek-V4.1-Flash-101717 Oct 2026
- Meta 11 Muse Spark Open15 Oct 2026
- xAI 14 Grok 4.721 Sep 2026
- Mistral 17 Mistral Large 4 (open weights)29 Oct 2026
- Alibaba 27 Qwen3.8-Max-09022 Sep 2026
- Moonshot 9 Kimi K3.1 Pro24 Oct 2026
- Zhipu 15 GLM-5.3-Flash26 Aug 2026
Each dot is one release at its date; the larger dots are landmark releases. The number beside a lab counts its frontier releases. Its name opens the lab’s profile, with a longer log of what it has shipped.
02 · The list
208 tracked
Every release, with what changed
Dates are to the month, checked in June and August 2026. The 100 rows with an exact day were each checked at the lab’s own page and link to it, and every lab’s release log has been read through 30 October 2026. Eight rows rest on press reports and say so.
Download the full list: CSV · JSON · free to reuse with credit (CC BY 4.0)
Showing 175 releases
| When | Lab | Model | What changed |
|---|---|---|---|
| 2026 | |||
| 29 Oct 2026 | Mistral | Mistral Large 4 (open weights) | Improved Mistral Large 4 after three more weeks of reinforcement learning, released as open weights under the Mistral Ecosystem Open-Weight licence; list price set at $0.70 in and $2.10 out Lab’s announcement |
| 26 Oct 2026 | OpenAI | GPT-6.1 Terra | Fast, efficient model for visual reasoning, computer use and writing, released in the API; rollout to ChatGPT Work and Codex for Plus and higher plans began the same day Lab’s announcement |
| 24 Oct 2026 | Moonshot | Kimi K3.1 Pro | Post-trained version of Kimi K3 with a 1M-token context, live on Kimi.com, Kimi Work, Kimi Code and the Kimi API on the day; open weights promised by 4 Nov Lab’s announcement |
| 17 Oct 2026 | DeepSeek | DeepSeek-V4.1-Flash-1017 | Re-post-trained V4.1-Flash checkpoint, live on the web, app and API and billed at the Flash price, with open weights Lab’s announcement |
| 15 Oct 2026 | Meta | Muse Spark Open | 816B-parameter (42B active) open-weight model on Muse Spark 1.3’s base, under the Muse Model License, with limited commercial terms Lab’s announcement |
| 13 Oct 2026 | Gemini 4 Argon (GA) | Generally available on the paid Gemini API and to Google AI Pro and Ultra subscribers in the Gemini app, after the restricted start on 30 Sep Lab’s announcement | |
| 12 Oct 2026 | Meta | Muse Wonder / Muse Wonder Full | First Muse Wonder models: Muse Wonder on the Meta Model API at $4 in and $20 out and in Muse Code; Muse Wonder Full for Muse Code subscribers only Lab’s announcement |
| 7 Oct 2026 | Anthropic | Claude Haiku 5.5not on the chart | Fast tier refresh, tuned for high-volume and latency-sensitive work Lab’s announcement |
| 6 Oct 2026 | Mistral | Mistral Large 4 | Flagship in public preview on the API from 6 Oct; an improved checkpoint followed as open weights on 29 Oct Lab’s announcement |
| 30 Sep 2026 | Gemini 4 Argonnot on the chart | Restricted at launch: trusted cyber defenders through Fairwind, plus early testers; generally available from 13 Oct (its own row) Lab’s announcement | |
| 29 Sep 2026 | OpenAI | GPT-6.1 Sol | Lower-cost professional and agentic model released in the API; rollout to ChatGPT Work and Codex began the same day Lab’s announcement |
| 28 Sep 2026 | Anthropic | Claude Sonnet 5.5 | Mid-tier model at $2 in and $10 out, on every platform from launch Lab’s announcement |
| 22 Sep 2026 | OpenAI | GPT-6 Sol / Luna | Two GPT-6 models released together Lab’s announcement |
| 22 Sep 2026 | Anthropic | Claude Opus 5.5 | Available on every platform from launch Lab’s announcement |
| 21 Sep 2026 | xAI | Grok 4.7 | Released to the API, Cursor and Grok Build Lab’s announcement |
| 10 Sep 2026 | DeepSeek | DeepSeek-V4.1-Flash | First model of a new architecture family, on the API Lab’s announcement |
| 4 Sep 2026 | OpenAI | GPT-6 Astra | First GPT-6 model; wide availability a day after a limited start Lab’s announcement |
| 2 Sep 2026 | Alibaba | Qwen3.8-Max-0902 | Upgraded snapshot of Qwen3.8-Max on the international and global API, aimed at long-horizon coding and multi-tool agent work; same price as Qwen3.8-Max Lab’s announcement |
| 2 Sep 2026 | Gemini 3.8 Flash | Flash tier, generally available from launch Lab’s announcement | |
| 2 Sep 2026 | Meta | Muse Spark 1.3 | In Muse Code and the Meta Model API (labelled public preview) from launch Lab’s announcement |
| 1 Sep 2026 | Anthropic | Claude Mythos 5.1not on the chart | Restricted sibling of Fable 5.1 for vetted cyberdefence and life-science organisations Lab’s announcement |
| 1 Sep 2026 | Anthropic | Claude Fable 5.1 | Generally available; its sibling Mythos 5.1 stayed partner-only Lab’s announcement |
| 26 Aug 2026 | Alibaba | Qwen3.8-Flashnot on the chart | Fast, low-cost Qwen3.8 model on the international and global API; the smallest of Alibaba’s three tiers Lab’s announcement |
| 26 Aug 2026 | Zhipu | GLM-5.3-Flash | First natively multimodal GLM-5 model (320B, 18B active) at about a tenth of GLM-5.2’s price; Z.ai’s second tier beside GLM-5.3, open weights under MIT Lab’s announcement |
| 18 Aug 2026 | Zhipu | GLM-5.3 | Successor to GLM-5.2: to GLM Coding Plan subscribers on 14 Aug, on the API from 18 Aug Lab’s announcement |
| 13 Aug 2026 | Gemini 3.7 Flash | Flash tier, generally available from launch Lab’s announcement | |
| 13 Aug 2026 | DeepSeek | DeepSeek-V4-Pro (GA) | General release of the V4-Pro that first shipped in April Lab’s announcement |
| 12 Aug 2026 | xAI | Grok 4.6 | Released in Cursor, Grok Build and the API, and through partners including OpenRouter, Vercel and Cloudflare Lab’s announcement |
| 10 Aug 2026 | Meta | Muse Glimmernot on the chart | 30B open-weight agent model under Apache 2.0 Lab’s announcement |
| 5 Aug 2026 | Meta | Muse Spark 1.2 | Opened to all developers through the Meta Model API (labelled public preview) and OpenRouter Lab’s announcement |
| 3 Aug 2026 | Alibaba | Qwen3.8-Max | Max tier, released through the API; open weights of a text-only version followed Lab’s announcement |
| 31 Jul 2026 | DeepSeek | DeepSeek-V4-Flash | Public beta of the Flash tier, re-trained from the April preview on the same architecture Lab’s announcement |
| 24 Jul 2026 | Anthropic | Claude Opus 5 | Opus-line generational: near-Fable-5 intelligence at half the price Lab’s announcement |
| 21 Jul 2026 | Gemini 3.6 Flash | Flash-tier refresh; Flash-Lite adds built-in computer use. No Pro-tier model shipped this cycle Lab’s announcement | |
| 19 Jul 2026 | Alibaba | Qwen3.8-Max Previewnot on the chart | 2.4T-parameter multimodal preview, only inside Alibaba’s coding and agent subscriptions (Token Plan, Qoder, QoderWork), not on the API; Qwen3.8-Max carries the dot Lab’s announcement |
| 16 Jul 2026 | Moonshot | Kimi K3 | 2.8T-parameter MoE with 1M context; the largest open-weight release to date, under a bespoke licence Lab’s announcement |
| 9 Jul 2026 | OpenAI | GPT-5.6 (Sol / Terra / Luna) | Three price-performance tiers, GA across ChatGPT, Codex and the API Lab’s announcement |
| 9 Jul 2026 | Meta | Muse Spark 1.1 | Multimodal reasoning for agentic work; 1M context; Meta Model API preview Lab’s announcement |
| 8 Jul 2026 | xAI | Grok 4.5 | Described by Musk as an Opus-class model; trained partly on real coding-session data Lab’s announcement |
| Jul 2026 | Mistral | Robostral Navigatenot on the chart | First embodied-navigation model |
| Jul 2026 | Meta | Muse Imagenot on the chart | First image model out of Meta Superintelligence Labs |
| Jul 2026 | Mistral | Leanstral 1.5not on the chart | Formal proof and verified coding |
| 30 Jun 2026 | Anthropic | Claude Sonnet 5 | Sonnet-line generational: close to Opus 4.8 at a fraction of the price Lab’s announcement |
| Jun 2026 | Mistral | Mistral OCR 4not on the chart | Document intelligence |
| Jun 2026 | Zhipu | GLM-5.2 | 753B-parameter MoE under an MIT licence, 1M context |
| Jun 2026 | Anthropic | Claude Fable 5 | First generally available Mythos-class model, with safety classifiers; 1M context. Mythos 5, the same model without them, stayed partner-only |
| Jun 2026 | Alibaba | Qwen3.7-Plus | Vision plus deep reasoning with tool invocation Dated from press reports |
| May 2026 | Anthropic | Claude Opus 4.8 | Flagship of the 4.x line |
| 19 May 2026 | Gemini 3.5 Flash | I/O launch: available to everyone in the Gemini app and AI Mode, and generally available in the Gemini API. Gemini 3.5 Pro was announced for the next month and never shipped Lab’s announcement | |
| May 2026 | Alibaba | Qwen3.7-Max | Proprietary max tier Dated from press reports |
| 5 May 2026 | OpenAI | GPT-5.5 Instant | ChatGPT’s new default for everyone, the tier just below GPT-5.5 Thinking; on the API as chat-latest Lab’s announcement |
| 28 Apr 2026 | Mistral | Mistral Medium 3.5 | Mid tier, powering remote coding agents; on the API from 28 Apr, launch post 22 May Lab’s announcement |
| Apr 2026 | OpenAI | GPT-5.5 | Flagship system of the 5.5 line |
| Apr 2026 | DeepSeek | DeepSeek-V4 (Pro/Flash) | Million-token context preview |
| 20 Apr 2026 | Moonshot | Kimi K2.6 | Agentic coding, general availability Lab’s announcement |
| Apr 2026 | Anthropic | Claude Opus 4.7 | Long-horizon agentic coding |
| Apr 2026 | xAI | Grok 4.3 Beta | Latest beta |
| 14 Apr 2026 | Alibaba | Qwen3.6-Max-Preview | Preview of the largest and most capable Qwen3.6 model, text only, on the international API Lab’s announcement |
| 8 Apr 2026 | Meta | Muse Spark | First closed-weight Meta frontier model, free in the Meta AI app and on meta.ai from launch, US first; the API was a private preview Lab’s announcement |
| 7 Apr 2026 | Zhipu | GLM-5.1 | Next-generation flagship for agentic engineering, open weights under MIT and on the API Lab’s announcement |
| Apr 2026 | Anthropic | Claude Mythos Previewnot on the chart | First Mythos-class model, a tier above Opus. Never generally released: critical-infrastructure partners only |
| 1 Apr 2026 | Alibaba | Qwen3.6-Plus | Plus tier of the Qwen3.6 vision-language series, on the international and global API; Qwen3.6-Flash and two small open models followed Lab’s announcement |
| Mar 2026 | OpenAI | GPT-5.4 | Point release |
| Mar 2026 | Mistral | Mistral Small 4not on the chart | Small tier refresh |
| 17 Mar 2026 | OpenAI | GPT-5.4 mininot on the chart | GPT-5.4’s mini, at $0.75 in and $4.50 out, released to the API with GPT-5.4 nano; third that day, below GPT-5.4 and GPT-5.3 Instant Lab’s announcement |
| 15 Mar 2026 | Zhipu | GLM-5-Turbo | Agent-tuned base model priced above GLM-5, on the general API Lab’s announcement |
| 3 Mar 2026 | OpenAI | GPT-5.3 Instant | ChatGPT’s new default for everyone, the tier just below Thinking; on the API as gpt-5.3-chat-latest Lab’s announcement |
| 17 Feb 2026 | Anthropic | Claude Sonnet 4.6 | Balanced tier upgrade; the default for Free and Pro on claude.ai, and on the API, from launch Lab’s announcement |
| Feb 2026 | Gemini 3.1 Pro | Preview update | |
| Feb 2026 | xAI | Grok 4.2 Beta | Beta release |
| Feb 2026 | Alibaba | Qwen3.5 | 397B mixture-of-experts flagship |
| Feb 2026 | Zhipu | GLM-5 | 744B MoE, first post-IPO flagship |
| Feb 2026 | OpenAI | GPT-5.3-Codex | Coding-focused Codex variant |
| Feb 2026 | Anthropic | Claude Opus 4.6 | Agent-team orchestration |
| 27 Jan 2026 | Moonshot | Kimi K2.5 | Multimodal update, on Kimi.com, the app, the API and Kimi Code from launch Lab’s announcement |
| 23 Jan 2026 | Alibaba | Qwen3-Max-Thinking | Max-tier reasoning model (qwen3-max-2026-01-23) in Qwen Chat and the paid API Lab’s announcement |
| 2025 | |||
| Dec 2025 | OpenAI | GPT-5.2 | Point release |
| 17 Dec 2025 | Gemini 3 Flash | New app default tier, rolled out to everyone at launch Lab’s announcement | |
| 22 Dec 2025 | Zhipu | GLM-4.7 | Year-end flagship Lab’s announcement |
| 1 Dec 2025 | DeepSeek | DeepSeek-V3.2 | Plus a Speciale variant Lab’s announcement |
| 2 Dec 2025 | Mistral | Mistral Large 3 | Mixture-of-experts flagship, open weights and on the API from launch Lab’s announcement |
| Nov 2025 | Anthropic | Claude Opus 4.5 | Flagship reasoning, frontier coding |
| 19 Nov 2025 | xAI | Grok 4.1 Fast | Fast tier of Grok 4.1 for agentic tool use, on the API Lab’s announcement |
| Nov 2025 | OpenAI | GPT-5.1 | Iterative quality and steerability upgrade |
| Nov 2025 | Gemini 3 Pro | Flagship launch, frontier reasoning | |
| Nov 2025 | xAI | Grok 4.1 | Broad rollout upgrade |
| Nov 2025 | Moonshot | Kimi K2 Thinking | Open reasoning and agentic model |
| 3 Nov 2025 | Alibaba | Qwen3-Max-Thinking Preview | Early preview of Qwen3-Max-Thinking, an intermediate checkpoint, in Qwen Chat and the API Lab’s announcement |
| Oct 2025 | Anthropic | Claude Haiku 4.5not on the chart | Fast tier refresh |
| 30 Sep 2025 | Zhipu | GLM-4.6 | Coding and agent upgrade |
| 29 Sep 2025 | Anthropic | Claude Sonnet 4.5 | Best agent value tier |
| 29 Sep 2025 | DeepSeek | DeepSeek-V3.2-Exp | Sparse-attention experiment, the default model on the app, web and API from launch Lab’s announcement |
| Sep 2025 | Alibaba | Qwen3-Max | Proprietary max tier of the Qwen3 family |
| 22 Sep 2025 | DeepSeek | DeepSeek-V3.1-Terminus | Final V3.1 update, on the app, web and API and as open weights Lab’s announcement |
| 19 Sep 2025 | xAI | Grok 4 Fast | Cost-efficient reasoning model at $0.20 in and $0.50 out, for all users including free and on the API; xAI’s second tier beside Grok 4 Lab’s announcement |
| 17 Sep 2025 | Mistral | Magistral Medium 1.2 | Multimodal update of the Medium-tier reasoning model, on the API Lab’s announcement |
| 11 Sep 2025 | Alibaba | Qwen3-Next | Ultra-sparse MoE architecture (80B, 3B active), open weights and on the API priced just above Plus Lab’s announcement |
| Sep 2025 | Moonshot | Kimi K2 Instruct 0905 | Longer context, stronger coding: a new K2 checkpoint, open weights and on the API |
| 5 Sep 2025 | Alibaba | Qwen3-Max-Preview | Preview of the trillion-parameter Max model in Qwen Chat and the paid API; the general release is its own row Lab’s announcement |
| Aug 2025 | Nano Banana (2.5 Flash Image)not on the chart | Viral image generation and editing | |
| Aug 2025 | DeepSeek | DeepSeek-V3.1 | Hybrid reasoning update |
| Aug 2025 | Zhipu | GLM-4.5Vnot on the chart | Vision-language model |
| 12 Aug 2025 | Mistral | Mistral Medium 3.1 | Medium-tier update on the API, improving tone and performance Lab’s announcement |
| Aug 2025 | OpenAI | GPT-5 | Unified system routing across model tiers |
| Aug 2025 | Anthropic | Claude Opus 4.1 | Coding and agent capability upgrade |
| 25 Jul 2025 | Alibaba | Qwen3-235B-A22B-Thinking-2507 | Named update of the open flagship (thinking), open weights and on the API Lab’s announcement |
| 24 Jul 2025 | Mistral | Magistral Medium 1.1 | Update of the Medium-tier reasoning model, on the API Lab’s announcement |
| 21 Jul 2025 | Alibaba | Qwen3-235B-A22B-Instruct-2507 | Named update of the open flagship (non-thinking), on the API between the Plus and Max prices Lab’s announcement |
| Jul 2025 | xAI | Grok 4 | Reasoning flagship, benchmark leader |
| Jul 2025 | Moonshot | Kimi K2 | Trillion-parameter MoE, open-sourced |
| 28 Jul 2025 | Zhipu | GLM-4.5 | Open agentic flagship, released with GLM-4.5-Air on the API and as open weights Lab’s announcement |
| 10 Jun 2025 | Mistral | Magistral | First reasoning models: Magistral Medium in public preview on the API and Le Chat, Magistral Small as open weights Lab’s announcement |
| May 2025 | DeepSeek | DeepSeek-R1-0528 | Major R1 upgrade |
| May 2025 | Anthropic | Claude Opus 4 / Sonnet 4 | Claude 4 generation, stronger agents |
| 7 May 2025 | Mistral | Mistral Medium 3 | Mid-tier flagship, on the API from launch Lab’s announcement |
| 6 May 2025 | Gemini 2.5 Pro (I/O edition) | Updated Gemini 2.5 Pro preview, released under its own name with its own post Lab’s announcement | |
| Apr 2025 | Alibaba | Qwen3 | Apache 2.0 family with hybrid thinking |
| Apr 2025 | OpenAI | o3 / o4-mini | Full reasoning lineup with tool use |
| Apr 2025 | OpenAI | GPT-4.1 | API-first coding and instruction model |
| 17 Apr 2025 | Gemini 2.5 Flash | Fast reasoning tier, in public preview in the Gemini API, AI Studio, Vertex AI and the Gemini app from 17 Apr; GA in June Lab’s announcement | |
| 14 Apr 2025 | Zhipu | GLM-4-Air-250414 | New Air-tier base and reasoning models (GLM-Z1-Air) on the API, with 32B open weights Lab’s announcement |
| Apr 2025 | Meta | Llama 4 (Scout/Maverick) | Mixture-of-experts, multimodal |
| Mar 2025 | Gemini 2.5 Pro | Thinking model, top of leaderboards | |
| 24 Mar 2025 | DeepSeek | DeepSeek-V3-0324 | Named V3 checkpoint behind deepseek-chat, open weights under MIT Lab’s announcement |
| 17 Mar 2025 | Mistral | Mistral Small 3.1 | Open multimodal Small update; Mistral’s second tier while no Medium was on sale Lab’s announcement |
| 5 Mar 2025 | Alibaba | QwQ-Plus | Reasoning model on the API priced between Plus and Max, with QwQ-32B as open weights Lab’s announcement |
| 24 Feb 2025 | Alibaba | QwQ-Max-Preview | Max-tier reasoning preview, free in Qwen Chat as “Thinking (QwQ)” Lab’s announcement |
| Feb 2025 | Anthropic | Claude Codenot on the chart | Agentic coding tool in the terminal |
| Feb 2025 | OpenAI | GPT-4.5 | "Orion": a large non-reasoning research preview |
| Feb 2025 | Anthropic | Claude 3.7 Sonnet | First hybrid reasoning model |
| Feb 2025 | xAI | Grok-3 | Colossus-trained flagship |
| 5 Feb 2025 | Gemini 2.0 Pro | Experimental public preview in AI Studio and Vertex AI, and in Gemini Advanced Lab’s announcement | |
| 30 Jan 2025 | Mistral | Mistral Small 3 | Efficient open model; Mistral’s second tier while no Medium was on sale Lab’s announcement |
| 28 Jan 2025 | Alibaba | Qwen2.5-Max | Large MoE flagship, on the API and Qwen Chat from launch Lab’s announcement |
| 25 Jan 2025 | Moonshot | Kimi K1.5 | Reasoning model matching o1, on the global Kimi web from 25 Jan; the API was by application Lab’s announcement |
| Jan 2025 | DeepSeek | DeepSeek-R1 | Open reasoning model that set off a global market shock |
| Jan 2025 | OpenAI | o3-mininot on the chart | Efficient reasoning at lower cost |
| 2024 | |||
| 19 Dec 2024 | Gemini 2.0 Flash Thinking | Experimental Flash-tier reasoning model, in public preview in AI Studio and the API Lab’s announcement | |
| Dec 2024 | Gemini 2.0 Flash | Agentic-era multimodal default | |
| Dec 2024 | DeepSeek | DeepSeek-V3 | 671B MoE trained at remarkably low cost |
| 10 Dec 2024 | DeepSeek | DeepSeek-V2.5-1210 | Named final V2.5 checkpoint, on the web and API and as open weights Lab’s announcement |
| Dec 2024 | OpenAI | o1 | Full release of the o1 reasoning model |
| Dec 2024 | Meta | Llama 3.3 | 70B efficiency update, the tier just below the 405B flagship |
| 20 Nov 2024 | DeepSeek | DeepSeek-R1-Lite-Preview | First DeepSeek reasoning model, free on chat.deepseek.com; no API or weights yet Lab’s announcement |
| 18 Nov 2024 | Mistral | Mistral Large 2.1 | Upgraded Large 2 on the API and Le Chat, open weights under the Mistral Research License Lab’s announcement |
| Nov 2024 | Alibaba | QwQ-32Bnot on the chart | Open reasoning model |
| 4 Nov 2024 | xAI | grok-beta | xAI’s API public beta opened with “a preview of a new Grok model” Lab’s announcement |
| Oct 2024 | Anthropic | Claude 3.5 Sonnet (upgraded) | Upgraded Sonnet, shipped with computer use in beta |
| Oct 2024 | Anthropic | Claude 3.5 Haikunot on the chart | Fast, low-cost tier refresh |
| Sep 2024 | Meta | Llama 3.2not on the chart | First multimodal Llama |
| Sep 2024 | Alibaba | Qwen2.5 | Broad size lineup, widely fine-tuned |
| Sep 2024 | OpenAI | o1-preview | First inference-time reasoning model |
| Sep 2024 | Mistral | Pixtralnot on the chart | First multimodal model |
| 5 Sep 2024 | DeepSeek | DeepSeek-V2.5 | Merged the chat and coder models; on the web and API and as open weights Lab’s announcement |
| Aug 2024 | Zhipu | GLM-4-Plus | Upgraded flagship Dated from press reports |
| Aug 2024 | xAI | Grok-2 | Added image generation |
| Jul 2024 | Meta | Llama 3.1 | First open model at frontier scale |
| Jul 2024 | Mistral | Mistral Large 2 | 123B-parameter flagship |
| 28 Jun 2024 | DeepSeek | DeepSeek-V2-0628 | Named V2 checkpoint behind deepseek-chat from 28 Jun, released as open weights (DeepSeek-V2-Chat-0628) on 18 Jul Lab’s announcement |
| Jun 2024 | Anthropic | Claude 3.5 Sonnet | Major coding leap, Artifacts launch |
| Jun 2024 | DeepSeek | DeepSeek-Coder-V2not on the chart | Strong open coding model |
| Jun 2024 | Alibaba | Qwen2 | Second generation: open weights up to 72B, the 72B served on Alibaba’s API at the Plus price |
| Jun 2024 | Zhipu | GLM-4-Air | The tier below GLM-4, close to its quality at lower cost, on the API |
| May 2024 | Mistral | Codestral 22Bnot on the chart | Dedicated code model |
| May 2024 | Alibaba | Tongyi Qianwen 2.5 (Qwen-Max-0428) | Flagship update behind Tongyi Qianwen 2.5, on the API and the Tongyi apps |
| May 2024 | OpenAI | GPT-4o | Omni real-time text, vision and audio |
| 14 May 2024 | Gemini 1.5 Flash | Fast, lightweight tier, in public preview from launch Lab’s announcement | |
| May 2024 | DeepSeek | DeepSeek-V2 | Cheap MoE that triggered China’s price war |
| 25 Apr 2024 | Alibaba | Qwen1.5-110B | Largest Qwen1.5 model, open weights and on the API above the Plus price Lab’s announcement |
| Apr 2024 | Meta | Llama 3 | 8B and 70B open models |
| Apr 2024 | Mistral | Mixtral 8x22Bnot on the chart | Larger open MoE |
| 9 Apr 2024 | Gemini 1.5 Pro | Million-token context, MoE architecture; a waitlisted preview from 15 Feb, public preview in the Gemini API from 9 Apr Lab’s announcement | |
| Apr 2024 | xAI | Grok-1.5 | Longer context and reasoning; announced 28 Mar for early testers and existing Grok users on X “in the coming days” Dated from press reports |
| Mar 2024 | Anthropic | Claude 3 (Opus/Sonnet/Haiku) | Three-tier family, first multimodal Claude |
| Feb 2024 | Mistral | Mistral Large | First proprietary flagship |
| 8 Feb 2024 | Gemini Ultra 1.0 | Google’s largest model, public in the paid Gemini Advanced plan as Bard became Gemini; announced in December, gated until then Lab’s announcement | |
| 4 Feb 2024 | Alibaba | Qwen1.5 | Open family up to 72B, on DashScope and together.ai Lab’s announcement |
| 16 Jan 2024 | Zhipu | GLM-4 | Flagship generation Lab’s announcement |
| Jan 2024 | Mistral | Mistral Medium | La Plateforme’s flagship endpoint: announced 11 Dec 2023 behind a waitlist, open to anyone who subscribed by 20 Jan 2024 |
| 2023 | |||
| Dec 2023 | Mistral | Mixtral 8x7B | Open mixture-of-experts breakthrough |
| Dec 2023 | Alibaba | Tongyi Qianwen 2.1 | Flagship update (qwen-max-1201) on the API |
| Dec 2023 | Gemini 1.0 | Gemini Pro in Bard and Nano on Pixel; Ultra followed in February | |
| Dec 2023 | xAI | Grok-1 | First Grok, launched inside X: a US waitlist from 3 Nov, open to all US Premium+ subscribers in December Dated from press reports |
| 30 Nov 2023 | Alibaba | Qwen-72B | Largest open Qwen model of its generation, open weights and on Alibaba’s API Lab’s announcement |
| Nov 2023 | DeepSeek | DeepSeek LLM | 67B base and chat models |
| Nov 2023 | OpenAI | GPT-4 Turbo | Cheaper, faster, longer context |
| Nov 2023 | Anthropic | Claude 2.1 | Doubled context, fewer refusals |
| Nov 2023 | Moonshot | Kimi (chatbot) | Long-context assistant: a closed beta from October, open to the public in November Dated from press reports |
| Nov 2023 | DeepSeek | DeepSeek Codernot on the chart | First model, code-specialised |
| 31 Oct 2023 | Alibaba | Tongyi Qianwen 2.0 | Second-generation flagship, public in China on the website and apps and through the API Lab’s announcement |
| Oct 2023 | Zhipu | ChatGLM3 | Third generation: a larger ChatGLM3 on the public chatglm.cn, and ChatGLM3-6B as open weights Dated from press reports |
| Sep 2023 | Mistral | Mistral 7B | First model, punched far above its size |
| Sep 2023 | Alibaba | Qwen (Tongyi Qianwen) | First public release: the flagship opened to the public on its website and to businesses through the API |
| Aug 2023 | Meta | Code Llamanot on the chart | Code-specialised variant |
| Aug 2023 | Zhipu | ChatGLM (chatglm.cn) | The larger ChatGLM flagship opened to the public on chatglm.cn after a closed beta since March Dated from press reports |
| Jul 2023 | Anthropic | Claude 2 | First public release, long context: claude.ai open to anyone in the US and UK |
| Jul 2023 | Meta | Llama 2 | First commercial open license |
| 10 May 2023 | PaLM 2 (Bard) | Bard moved to PaLM 2 and opened to everyone in more than 180 countries, its waitlist removed Lab’s announcement | |
| Mar 2023 | OpenAI | GPT-4 | Multimodal flagship, large capability jump |
| Mar 2023 | Anthropic | Claude 1 | First Claude, Constitutional AI approach; the API was by request, the free Claude app for Slack open to any team from 30 Mar |
| Mar 2023 | Zhipu | ChatGLM-6Bnot on the chart | First open GLM chat model |
| Feb 2023 | Meta | LLaMA 1not on the chart | Research weights that leaked and seeded the open scene |
| 2022 | |||
| Nov 2022 | OpenAI | text-davinci-003 | OpenAI’s most capable model on its API, released two days before ChatGPT |
| Nov 2022 | OpenAI | ChatGPT | Consumer chatbot; the fastest app to reach 100M users |
| Mar 2022 | OpenAI | GPT-3.5not on the chart | InstructGPT-tuned for following instructions |
| 2020 | |||
| Jun 2020 | OpenAI | GPT-3not on the chart | 175B-parameter model behind a public API |
| 2019 | |||
| Feb 2019 | OpenAI | GPT-2not on the chart | Staged release over 2019 on safety concerns |
| 2018 | |||
| Jun 2018 | OpenAI | GPT-1not on the chart | First generative pretrained transformer |
03 · Chart No. 1
the argument, drawn by hand
The argument, drawn on the same dates
Each dot is a release from the list above, at its real date. Its height is not a measurement. Every dot sits on the grey curve, and both curves are this publication’s argument, drawn by hand.
The grey curve is how much of its margin a newly released model keeps: it peaks with GPT-4 and falls as each new model is matched and priced down. The gold curve is the value of what the models run on and plug into, which rises as they get cheaper. The crossing is a trend in the drawing, with no date attached.
The case is made in The Model Is Not the Moat, The Forest Floor Is the Product and The Five Laws of Durable Systems.
The newsletter · Free
Subscribe to get each new piece in your inbox, with something to use in every one.
Several essays a week, plus Notes through the week.
Signup completes on Substack. Unsubscribe from any email.