Model releases

Every notable model since ChatGPT, newest first: what shipped, when, and what it changed.

202631 releases

15 Oct

Muse Spark Open Meta

Open-weights version of Muse Spark; 816B mixture-of-experts with 42B active, 1M context, text, image, video and audio input.

1M context
Open weights

Current

15 Oct

Nemotron 3.5 Super VL Nvidia

New 120B (12B active) vision-language model in the Nemotron 3.5 family; up to 1M context, image and video input, open datasets and recipes.

1M context
Open weights

Current

15 Oct

Nemotron 3.5 Bolt Nvidia

New local tier of the Nemotron 3.5 family: 74B total, 9B active, 2-bit native quantisation to fit in 24GB.

256K context
Open weights
Free (open weights) / Free (open weights) per M tokens checked 19 Oct 2026

Current

13 Oct

Gemini 4 Argon Google

Google's latest Gemini model; 1M unified input and output limit, 64K default output (up to 1M with long decode), low, medium and high thinking levels.

1M context
Closed weights
$2.00 / $10.00 per M tokens checked 16 Oct 2026

Current

12 Oct

Muse Wonder Meta

New flagship above Muse Spark; 1M context, text, image, video, audio and PDF input; sibling Muse Wonder Full is subscription-only in Muse Code.

1M context
Closed weights

Preview

7 Oct

Claude Haiku 5.5 Anthropic

Succeeds Haiku 4.5; 1M context, 128K max output, adaptive thinking.

1M context
Closed weights
$0.10 / $0.50 per M tokens checked 12 Oct 2026

Current

6 Oct

Mistral Large 4 Mistral AI

Succeeds Mistral Large 3 with a 1T-parameter (52B active) multimodal MoE, a 1M context and built-in reasoning.

1M context
Closed (open weights promised; see footnotes) weights

Preview

29 Sep

GPT-6.1 Sol OpenAI

Succeeds GPT-6 Sol at the same price; cached input halved to $0.10/M.

1.05M context
Closed weights
$2 / $10 per M tokens checked 12 Oct 2026

Current

28 Sep

Claude Sonnet 5.5 Anthropic

Succeeds Sonnet 5 at the same price; 1M context, 128K max output.

1M context
Closed weights
$2 / $10 per M tokens checked 28 Sept 2026

Current

22 Sep

GPT-6 Sol OpenAI

Mid-tier GPT-6 model; superseded by GPT-6.1 Sol.

Closed weights

Superseded

22 Sep

GPT-6 Luna OpenAI

Low-cost GPT-6 model.

Closed weights
$0.10 / $0.50 per M tokens checked 12 Oct 2026

Current

21 Sep

Grok 4.7 xAI

Succeeds Grok 4.6 with a new, larger base model, a 500K context and four reasoning effort levels, at the same price.

500K context
Closed weights
$2.00 / $6.00 per M tokens checked 19 Oct 2026

Current

10 Sep

DeepSeek-V4.1-Flash DeepSeek

Succeeds V4-Flash (retired) on a new encoder-decoder architecture: 552B total, 8B/16B active, 384K max output, thinking on by default, native vision. Refreshed as V4.1-Flash-1017 on 17 Oct 2026.

1M context
Open weights
$0.30 / $1.20 per M tokens checked 20 Oct 2026

Current

3 Sep

GPT-6 Astra OpenAI

OpenAI's new flagship ("our most capable model"), above GPT-5.6.

1.05M context
Closed weights
$10 / $50 per M tokens checked 12 Oct 2026

Current

2 Sep

Gemini 3.8 Flash Google DeepMind

Near-frontier reasoning at Flash speed, with an adjustable effort level for tool calling.

1M context
Closed weights
$0.75 / $3.75 per M tokens checked 12 Oct 2026

Current

2 Sep

Muse Spark 1.3 Meta

Succeeds Muse Spark 1.2; max reasoning mode, better long-horizon agent work, about 20% fewer tool calls and 25% fewer tokens in coding.

1M context
Closed weights
$1.25 / $4.25 per M tokens checked 18 Oct 2026

Current

18 Aug

GLM-5.3 Z.ai

Went to a million-token context and made reasoning the default rather than an option.

1M context
Open weights

Current

12 Aug

Grok 4.6 xAI

Retrained for long-running agentic and visual work. Superseded by Grok 4.7 (21 Sep 2026).

Closed weights

Superseded

12 Aug

MAI-Thinking-1 Microsoft

Microsoft’s first reasoning model, trained without distilling from third parties.

256K context
Closed weights

Current

3 Aug

Qwen3.8-Max Alibaba

Alibaba's new flagship after the Qwen3 line: 2.4T total / 95B active MoE, 1M context, 131K max output, thinking mode, image and video input.

1M context
Open weights
$2.00 / $6.00 per M tokens checked 20 Oct 2026

Current

24 Jul

Claude Opus 5 Anthropic

Set a new bar on coding and knowledge work at roughly half the price of the tier above it.

1M context
Closed weights
$5 / $25 per M tokens checked 12 Oct 2026

Current

17 Jul

Kimi K3 Moonshot AI

A 2.8-trillion-parameter model that added native video understanding to the Kimi line.

1M context
Open weights

Current

9 Jul

GPT-5.6 OpenAI

Split the flagship into three tiers and added a mode that runs parallel agent workstreams.

1M context
Closed weights

Current

9 Jul

Muse Spark 1.1 Meta

Added a public Meta Model API and a thinking mode, with a million-token context. Superseded by Muse Spark 1.2 and then Muse Spark 1.3 (2 Sep 2026).

1M context
Closed weights

Superseded

30 Jun

Claude Sonnet 5 Anthropic

Brought near-Opus results down to Sonnet pricing, with a million-token context.

1M context
Closed weights

Superseded

19 May

Gemini 3.5 Flash Google DeepMind

Turned the workhorse line toward long-horizon agentic tasks rather than raw benchmark scores.

1M context
Closed weights

Deprecated

24 Apr

DeepSeek-V4 DeepSeek

Brought a million-token context and sparse attention to the open-weight world.

1M context
Open weights

Current

23 Apr

GPT-5.5 OpenAI

A large agentic jump for long-running coding and research work without the usual latency cost.

1M context
Closed weights

Superseded

8 Apr

Muse Spark Meta

Meta dropped open weights for its flagship for a closed model from the new Superintelligence Labs.

Closed weights

Superseded

2 Apr

Gemma 4 Google DeepMind

Brought native audio input and agentic function calling to Google’s open weights.

256K context
Open weights

Current

19 Feb

Gemini 3.1 Pro Google DeepMind

More than doubled its ARC-AGI-2 reasoning score against Gemini 3 Pro.

1M context
Closed weights
$2 / $12 per M tokens checked 12 Oct 2026

Current

202535 releases

17 Dec

Gemini 3 Flash Google DeepMind

Three times faster than the previous generation while beating 2.5 Pro on coding.

1M context
Closed weights

Superseded

15 Dec

Nemotron 3 Nvidia

Hybrid latent mixture-of-experts; Nano gained four times the throughput and a 1M context.

1M context
Open weights

Current

11 Dec

GPT-5.2 OpenAI

Tripled its score on the ARC-AGI-2 reasoning benchmark and halved chart and interface vision errors.

256K context
Closed weights

Superseded

2 Dec

Mistral Large 3 Mistral AI

Return to mixture-of-experts at 675B total parameters, released under Apache 2.0.

256K context
Open weights
$2 / $6 per M tokens checked 12 Sept 2026

Current

2 Dec

Amazon Nova 2 Amazon

Added three levels of extended reasoning plus a built-in code interpreter.

1M context
Closed weights

Current

24 Nov

Claude Opus 4.5 Anthropic

Added automatic context summarisation, so long conversations stop hitting a wall.

200K context
Closed weights

Superseded

18 Nov

Gemini 3 Pro Google DeepMind

Folded multimodality, thinking and tool use into a single flagship with agentic coding.

1M context
Closed weights

Superseded

13 Nov

GPT-5.1 OpenAI

Made thinking: it spends effort in proportion to how hard the question is.

400K context
Closed weights

Superseded

6 Nov

Kimi K2 Thinking Moonshot AI

Sustained two to three hundred sequential tool calls without losing the thread.

256K context
Open weights

Superseded

15 Oct

Claude Haiku 4.5 Anthropic

Matched the previous Sonnet on coding at a third of the cost and twice the speed.

200K context
Closed weights

Superseded

8 Oct

Jamba Reasoning 3B AI21

A 3B reasoning model for on-device use, with an eight-times smaller KV cache.

256K context
Open weights

Current

30 Sep

GLM-4.6 Z.ai

Extended context to 200K and sharpened real-world coding performance.

200K context
Open weights

Superseded

29 Sep

Claude Sonnet 4.5 Anthropic

Held focus on autonomous work for 30 hours and more, a step change in agent endurance.

200K context
Closed weights

Deprecated

28 Aug

MAI-1-preview Microsoft

Microsoft’s first in-house foundation models, ending sole reliance on OpenAI.

Closed weights

Superseded

20 Aug

Seed-OSS-36B ByteDance

ByteDance’s first broadly open model, with a 512K context and an adjustable thinking budget.

512K context
Open weights

Current

7 Aug

GPT-5 OpenAI

Merged the chat and reasoning lines into one model with a dial for how long it thinks.

400K context
Closed weights

Superseded

5 Aug

gpt-oss-120b and gpt-oss-20b OpenAI

OpenAI’s first open-weight release since GPT-2, Apache-licensed and runnable on one GPU.

128K context
Open weights

Current

5 Aug

Claude Opus 4.1 Anthropic

A coding and agentic-research refresh, taking SWE-bench Verified to 74.5%.

200K context
Closed weights

Retired

28 Jul

GLM-4.5 Z.ai

MIT-licensed 355B agentic model, among the most permissive frontier-scale releases.

128K context
Open weights

Superseded

11 Jul

Kimi K2 Moonshot AI

A trillion-parameter open mixture-of-experts model built specifically for agentic tool use.

128K context
Open weights

Superseded

9 Jul

Grok-4 xAI

Native tool use, with reinforcement learning applied at pretraining scale.

256K context
Closed weights

Superseded

16 Jun

MiniMax-M1 MiniMax

First open large-scale hybrid-attention reasoning model, at a third of R1’s compute.

1M context
Open weights

Superseded

22 May

Claude Opus 4 and Sonnet 4 Anthropic

Added tool use during extended thinking, plus memory files that persist across sessions.

200K context
Closed weights

Retired

29 Apr

Qwen3 Alibaba

Hybrid thinking and non-thinking modes in one model, trained across 119 languages.

128K context
Open weights

Superseded

16 Apr

o3 and o4-mini OpenAI

First reasoning models that agentically use every tool (browsing, code, files) and reason over images.

200K context
Closed weights

Superseded

14 Apr

GPT-4.1 OpenAI

Took context to a million tokens and posted a large coding gain over GPT-4o.

1M context
Closed weights

Superseded

5 Apr

Llama 4 Meta

First natively multimodal Llama, mixture-of-experts, with a ten-million-token context on Scout.

10M context
Open weights

Superseded

25 Mar

Gemini 2.5 Pro Google DeepMind

Google’s first. It reasons through steps before it answers.

1M context
Closed weights

Superseded

13 Mar

Command A Cohere

Matched GPT-4o on agentic tasks while running on just two GPUs.

256K context
Open weights

Current

12 Mar

Gemma 3 Google DeepMind

First Gemma with vision and a 128K context, covering 140-plus languages.

128K context
Open weights

Superseded

27 Feb

GPT-4.5 OpenAI

The largest non-reasoning chat model OpenAI shipped; big factuality gain, short commercial life.

128K context
Closed weights

Retired

24 Feb

Claude 3.7 Sonnet Anthropic

First hybrid model with extended thinking you can switch on; Claude Code shipped alongside it.

200K context
Closed weights

Retired

19 Feb

Grok-3 xAI

Eight times the context, a Think reasoning mode and the DeepSearch agent.

1M context
Closed weights

Superseded

20 Jan

DeepSeek-R1 DeepSeek

First open-weight reasoning model to match o1, the release that moved markets.

128K context
Open weights

Retired

15 Jan

MiniMax-01 MiniMax

A four-million-token context via lightning attention, far beyond anything else open at the time.

4M context
Open weights

Superseded

202426 releases

26 Dec

DeepSeek-V3 DeepSeek

Reached GPT-4o-class benchmarks on a reported training budget under six million dollars.

128K context
Open weights

Superseded

11 Dec

Gemini 2.0 Flash Google DeepMind

Added native image and audio output plus built-in tool use, aimed squarely at agents.

1M context
Closed weights

Superseded

6 Dec

Llama 3.3 70B Meta

Matched the 405B model’s benchmarks at a sixth of the size, through post-training alone.

128K context
Open weights

Superseded

3 Dec

Amazon Nova Amazon

Replaced Titan with a multimodal family spanning text, image and video.

300K context
Closed weights

Superseded

22 Oct

Claude 3.5 Sonnet with computer use Anthropic

First frontier model given direct control of a desktop, moving the cursor and clicking.

200K context
Closed weights

Retired

25 Sep

Llama 3.2 Meta

Added vision at 11B and 90B, plus 1B and 3B models small enough for phones.

128K context
Open weights

Superseded

19 Sep

Qwen2.5 Alibaba

A very wide release: 0.5B to 72B, plus dedicated coding and maths models.

128K context
Open weights

Superseded

12 Sep

o1-preview OpenAI

The first. It thinks in a hidden chain of thought before answering.

128K context
Closed weights

Retired

13 Aug

Grok-2 xAI

Added image understanding and image generation.

Closed weights

Superseded

24 Jul

Mistral Large 2 Mistral AI

123B parameters and a 128K context, closing much of the gap to GPT-4o.

128K context
Open weights

Superseded

23 Jul

Llama 3.1 405B Meta

First open-weight model Meta claimed was competitive with closed frontier models.

128K context
Open weights

Superseded

18 Jul

GPT-4o mini OpenAI

Replaced GPT-3.5 Turbo as the cheap default, at close to GPT-4o quality.

128K context
Closed weights

Superseded

27 Jun

Gemma 2 Google DeepMind

Its 27B model matched systems twice its size and ran on a single host.

8K context
Open weights

Superseded

20 Jun

Claude 3.5 Sonnet Anthropic

A mid-tier model that beat the previous generation’s largest model on most benchmarks.

200K context
Closed weights

Retired

14 Jun

Nemotron-4 340B Nvidia

An open family built to generate synthetic training data for other people’s models.

Open weights

Superseded

7 Jun

Qwen2 Alibaba

Added grouped-query attention across all sizes and training in 27 more languages.

128K context
Open weights

Superseded

13 May

GPT-4o OpenAI

First natively multimodal model: one network for text, audio and vision, with near-instant voice.

128K context
Closed weights

Retired

7 May

DeepSeek-V2 DeepSeek

A 236B mixture-of-experts model trained at a fraction of Western cost, and priced accordingly.

128K context
Open weights

Superseded

23 Apr

Phi-3 Microsoft

Showed a small model trained on carefully curated data could punch far above its size.

128K context
Open weights

Superseded

18 Apr

Llama 3 Meta

New tokeniser and architecture at 8B and 70B, with a 400B model still training.

8K context
Open weights

Superseded

28 Mar

Jamba AI21

First production-grade hybrid of Mamba state-space layers and transformers.

256K context
Open weights

Superseded

24 Mar

Command R Cohere

Built for retrieval-augmented generation at production scale rather than for chat.

128K context
Open weights

Superseded

4 Mar

Claude 3 family Anthropic

Introduced the Haiku, Sonnet and Opus tiers and gave Claude vision for the first time.

200K context
Closed weights

Retired

26 Feb

Mistral Large Mistral AI

Mistral’s first closed flagship, sold through an API rather than given away.

32K context
Closed weights

Superseded

21 Feb

Gemma Google DeepMind

Google’s first open-weight models, built from the same research as Gemini.

8K context
Open weights

Superseded

15 Feb

Gemini 1.5 Pro Google DeepMind

Jumped to a million-token context using a mixture-of-experts design, an order-of-magnitude leap.

1M context
Closed weights

Superseded

202313 releases

11 Dec

Mixtral 8x7B Mistral AI

First open mixture-of-experts model: 47B total, only 13B active per token.

32K context
Open weights

Superseded

6 Dec

Gemini 1.0 Google DeepMind

First model pre-trained multimodally from the start rather than bolting vision onto a text model.

32K context
Closed weights

Superseded

21 Nov

Claude 2.1 Anthropic

Doubled context to 200K, cut hallucination rates and added tool use in beta.

200K context
Closed weights

Retired

6 Nov

GPT-4 Turbo OpenAI

Added vision input and pushed context from 8K to 128K at roughly a third of the price.

128K context
Closed weights

Superseded

3 Nov

Grok-1 xAI

xAI’s first model, wired to live X data, in closed beta.

Closed weights

Retired

27 Sep

Mistral 7B Mistral AI

A seven-billion-parameter debut that beat 13B rivals, under a permissive Apache licence.

4K context
Open weights

Superseded

18 Jul

Llama 2 Meta

First Llama licensed for commercial use, which made open weights a business option.

4K context
Open weights

Superseded

11 Jul

Claude 2 Anthropic

First Claude anyone could sign up for, with a 100K context window when rivals offered 8K.

100K context
Closed weights

Retired

10 May

PaLM 2 Google

A four-size family with much stronger multilingual, reasoning and coding ability than LaMDA.

Closed weights

Retired

21 Mar

Bard Google

Google’s answer to ChatGPT, built on a lightweight LaMDA and later folded into Gemini.

Closed weights

Retired

14 Mar

GPT-4 OpenAI

A large jump in accuracy and reliability over GPT-3.5, and the first to pass professional exams at scale.

8K context
Closed weights

Retired

14 Mar

Claude 1 Anthropic

Anthropic’s first public model, launched the same day as GPT-4.

9K context
Closed weights

Retired

24 Feb

Llama 1 Meta

Meta’s first foundation family,, research-licensed and promptly leaked, seeding the open-model scene.

2K context
Open weights

Retired

20221 release

30 Nov

ChatGPT (GPT-3.5) OpenAI

The first public chat product: GPT-3 tuned with RLHF for conversation rather than raw completion.

4K context
Closed weights

Retired

Dates are the public announcement date, which sometimes ran ahead of general availability. Model names link to the vendor’s own announcement. Prices, where shown, are per million tokens and carry the date we last checked them.