AI model release archive

A more heavily post-trained Kimi K3 that cuts its claim-level hallucination rate from 23.4% to 8.3% and keeps the 1M-token context, with open weights promised by November 4.

An update to DeepSeek-V4.1-Flash that scores higher on every agentic, coding and reasoning test DeepSeek published while giving shorter answers, with the same MIT weights, 1M-token context and API price.

Brings FLUX 3 Image's editing, up to ten reference images and bounding-box layout to a 5.6B open-weights model that runs on consumer GPUs down to an RTX 3090, with output up to 4K.

Meta's first downloadable Muse Spark: an 816B-parameter mixture-of-experts model with about 42B active parameters, a 1M-token context and text, image, video and audio input, released under the Muse Model License.

An open-weight model with 120B parameters, 12B active at a time, that reads text, images and videos up to 2 minutes long, with a context of up to 1M tokens.

Doubles the longest clip to 30 seconds from Kling 3.0's 15, outputs two-channel stereo sound, and accepts up to 10 keyframe images and prompts of up to 8,000 tokens.

Google's first Gemini 4 model replaces Gemini 3.1 Pro as its top model in the paid Gemini API, with a 1M-token window shared by input, thinking and output.

Meta's first frontier-tier model, a larger step above Muse Spark built for agents, scoring 4 points higher on the Artificial Analysis index at about three to five times the API price.

Costs a tenth of Claude Haiku 4.5 for prompts under 100K tokens while raising its computer-use score from 15.7% to 72.4%, with a 1M-token context window.

An update to Nano Banana 2 built on Gemini 3.6 Flash that scores 60 Elo higher in Google's blind text-to-image tests at about half the price per image.

Mistral's first large reasoning model, natively multimodal with 1 trillion parameters (49 billion active), lifts the Intelligence Index from 9 to 38 but costs almost three times as much as Mistral Large 3.

Generates images natively at 2K and 4K and adds bounding-box layout control to place, re-describe or replace individual elements.

One week after GPT-6 Sol, it scores 7 points higher on computer-use tasks and 4 points higher on the all-round intelligence index at the same $2 / $10 price.

Scores 18 points higher than Sonnet 5 on the all-round intelligence index and runs over 30% faster, at the same $2 / $10 price per 1M tokens.

MiniMax's first lighter, closed Flash model; generally available in the API since October 19, 2026 at $0.15 / $0.60 per 1M tokens.

Half the price of GPT-5.6 Sol with about half as many factual mistakes, while its all-round score rises only one point.

Half the input price and 58% less on output than GPT-5.6 Luna, with a 1-point higher all-round score.

Anthropic's new Opus scores 7 points higher than Opus 5 on the all-round index and costs 20% less per token.

Xiaomi's new open-weights flagship scores 46 on the Artificial Analysis Intelligence Index, up from 26 for MiMo-V2.5-Pro, at the same API price of $0.435 / $0.87 per 1M tokens.

Xiaomi's new open-weights Flash model beats its previous Pro model, MiMo-V2.5-Pro, on all 14 tests where Xiaomi's table scores both, while keeping MiMo-V2.5's price of $0.14 / $0.28 per 1M tokens.

A larger base model than Grok 4.6, trained longer on multi-hour agentic tasks, at the same $2 / $6 price and the same 500k-token context.

A 7B open-weight model that unifies generation and editing, adds native transparent (RGBA) output and up to 10 reference images, under a research-only license.

Qwen's new omni-modal model for agent work: it takes text, images, audio and video, replies in text, and has a 1M-token context window, up from 256K on Qwen3.5-Omni.

Kimi K2.8 Preview rolled out in Kimi Code with Kimi K3's low / high / max thinking levels and a context window of up to 1M tokens on every membership tier, up from 262,144 tokens for Kimi K2.7 Code.

A new encoder-decoder design that activates 8B parameters on input and 16B on output, beating V4-Flash-0731 on every agentic test DeepSeek published, with the MIT licence, a 1M-token context and a lower API price than the model it replaces.

Ships as two variants in ChatGPT and the API: Flare, with higher quality than GPT Image 2 at 50% lower latency, and Sunburst, a slower model for precise editing work.

OpenAI's new flagship scores 6 points higher than GPT-5.6 Sol on the all-round intelligence index, makes up far fewer answers and uses a third of the tokens on coding-agent work — at 2.5 times the price.

A faster version of MAI-Image-2.6 that costs about half as much per image and scores 47 Elo lower in Artificial Analysis blind image comparisons.

Scores 89.4% on Terminal-Bench 2.1, up from 85.8% for Gemini 3.7 Flash, at the same introductory price as Gemini 3.7 Flash.

Five points up on the all-round intelligence index over Muse Spark 1.2 and higher scores on office and terminal tasks, at the same price and the same 1M-token context.

An upgraded snapshot of Qwen 3.8-Max, further trained for coding and office-style agent work, with the same 2.4 trillion parameters, 1M-token context window and $2 / $6 price.

Anthropic's top model, refreshed: big jumps on long coding and scientific-research tasks, and cached input now costs 75% less — same headline price as Fable 5.

Scene extension now reads 10 seconds of prior footage (up from 1) for clips up to 40 seconds, adds first-and-last-frame camera control, 4K upscaling and a 360p draft mode that renders up to 60% faster at a third of the cost.

Open-weight mixture-of-experts model with 125B total parameters but only 6B active per token, a 262K native context window and a built-in vision encoder, previewing the Qwen4 architecture.

GLM-5.3-Flash is a 320B-parameter open model (18B active) with vision and a 1M-token context window that Z.ai compares with GLM-5.2: it solves 63.4% of DeepSWE v1.1 coding tasks versus 46.2%, at about a tenth of the API price.

Doubles the maximum clip length to 30 seconds in a single pass, accepts documents (PDF, PPT, DOC) as input alongside text, image, audio and video, and improves face realism and reference consistency.

Microsoft's updated image model scores 46 Elo points above MAI-Image-2.5 in Artificial Analysis blind comparisons and costs about a fifth less per image; it is in Public Preview on Microsoft Foundry.

Alibaba's open 27B model scores 84.3% on OSWorld-Verified, up from 63.9% for Qwen3.6-27B, and stays free to download under Apache 2.0.

GLM-5.3 keeps GLM-5.2's base model and API price but, after more post-training, completes 28.3% of Terminal-Bench 3.0 tasks instead of 4.6% and 66.9% of DeepSWE v1.1 tasks instead of 46.2%.

Five points up on the all-round intelligence index over 3.6 Flash and about 40% faster per task, at a discounted $0.75 / $3.75 per 1M tokens through the end of 2026.

The general-availability release of V4-Pro scores 87.9% on Terminal Bench 2.1, up from 72.1% for the April preview in DeepSeek's tests, and keeps the MIT license.

Builds on Grok 4.5 with a longer extra training run; adds an xhigh reasoning-effort level, and cached input now costs $0.50 per million tokens, up from $0.30.

Open-weights release of Qwen 3.8-Max, without the cloud model's image input and non-thinking mode; a commercial licence is required above $50M in annual revenue.

Microsoft AI's first in-house reasoning model, open to everyone as a public preview in Microsoft Foundry.

Adds image input and a 256K-token context, solves more terminal tasks with fewer tokens than MAI-Code-1-Flash, and Microsoft says it costs a quarter as much to run — available to every GitHub Copilot user.

Meta's 30B open-weight model under Apache 2.0 (its most permissive license yet), built to run AI agents locally; quantized, it fits on a single 24–32 GB GPU.

xAI's new image model, built for precise editing, sharp text rendering and consistent multi-image workflows.

A coding-focused update to Muse Spark, released with Meta's Muse Code agent at the same $1.25 / $4.25 price, plus a cheaper contributor tier ($0.10 / $0.20 per million tokens) in which Meta may train on your prompts and completions.

The official release of V4-Flash keeps the preview's 284B-parameter design but, after further training, scores 82.7% on Terminal Bench 2.1, up from 61.8%, in DeepSeek's tests.

Priced at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8, and became the default model for Claude Max subscribers.

Released alongside 3.5 Flash-Lite; Google reported up to 17% fewer tokens used than the previous Flash model.

The lowest-cost model of Google's Flash line, released the same day as 3.6 Flash.

Alibaba's third-generation Qwen-Image model, available only through its API and apps with no open weights, taking prompts up to 4.5k tokens, rendering text as small as 10 pixels and supporting 12 languages.

Alibaba's largest Qwen model so far, with 2.4 trillion parameters (about 95 billion active per token) and a 1M-token context window, priced 20% below Qwen 3.7-Max.

Kimi K3 replaces Kimi K2.6 with a 2.8-trillion-parameter open model whose context window is four times larger (1M tokens) and whose API price is more than three times higher.

The most capable of the three GPT-5.6 models (Sol, Terra and Luna), priced at $5 per million input tokens and $30 per million output tokens.

The mid tier of the GPT-5.6 trio (Sol / Terra / Luna), positioned by OpenAI as the balanced everyday model: $2 per million input tokens and $12 per million output tokens after the 30 July price cut.

The lightest tier of the GPT-5.6 trio, positioned by OpenAI as the fast, affordable model: $0.20 per million input tokens and $1.20 per million output tokens after the 30 July price cut.

Opened the Meta Model API to outside developers in public preview, priced at $1.25 per million input tokens and $4.25 per million output tokens.

Added configurable reasoning effort (low, medium, high) with a 500K-token context window, priced at $2 per million input tokens and $6 per million output tokens.

Meta's first in-house image model, free in the Meta AI app, Instagram Stories and WhatsApp.

It replaces GPT-5.3 Instant Mini as the fallback model ChatGPT users get after reaching their GPT-5.5 Instant or Auto limits; it does not appear in the model picker or the API.

Google's fastest and cheapest image model: text-to-image in about 4 seconds at $0.034 per 1K image, half the price of Nano Banana 2, built for high-volume generation.

Replaced Sonnet 4.6 as the default model on Anthropic's Free and Pro plans, priced at $2 per million input tokens and $10 per million output tokens.

Apache 2.0 model for Lean 4 proof engineering with 119B parameters and 6B active, solving 587 of 672 PutnamBench problems.

Alibaba's updated video model scores higher than HappyHorse 1.0 in Artificial Analysis blind comparisons and costs 25% less for 1080p video, $10.80 instead of $14.40 per minute.

Google's first open-weight text diffusion model: 26B parameters with 3.8B active, Apache 2.0, generating over 1,000 tokens per second on a single H100.

Anthropic's new top-end model scores 80.3% on SWE-bench Pro (Anthropic gives one figure for Fable 5 and Mythos 5), up from 69.2% for Opus 4.8, at twice the price.

xAI's video model that turns a still image into a clip with sound effects, ambience and dialogue generated in the same pass; a July 31 update added text-to-video, reference-to-video and native 1080p.

Microsoft's first in-house coding model — small, cheap, and ahead of Claude Haiku 4.5 on the coding tests Microsoft published.

Launched at No. 2 for image editing on Arena, unveiled at Build 2026.

Alibaba's closed multimodal agent model with a 1M-token context, offered through Alibaba Cloud Model Studio at $0.40 per 1M input and $1.60 per 1M output tokens (international pricing, prompts up to 256K tokens).

Scores 74.6% on Terminal-Bench 2.1, up from 66.1% for Opus 4.7, at the same $5 / $25 price per 1M tokens, and its fast mode now costs a third of what fast mode cost for earlier models.

The largest model in Alibaba's Qwen3.7 series, built for coding, office work and long autonomous tasks, with a 1M-token context window, up from 262K on Qwen3.6-Max-Preview.

xAI's coding model behind its Grok Build agent took over from Grok Code Fast 1 on the API, served at over 100 tokens per second and priced at $1 / $2 per million tokens, up from $0.20 / $1.50.

Google's lightweight Flash model: it edges out the bigger Gemini 3.1 Pro on a coding test, while running far more token-efficient than the Flash it replaces — at the same price.

Google's new any-to-video model: text, images, audio or video in, physics-aware video out, editable by chatting — 10-second clips, free in YouTube Shorts and included with Google AI Plus, Pro and Ultra in the Gemini app.

OpenAI's new lightweight model: far fewer made-up answers and much stronger math than GPT-5.3 Instant Mini — and it's now the default free model in ChatGPT.

Mistral Medium 3.5 merges instruction-following, reasoning and coding into one 128B dense model released with open weights, with configurable reasoning effort and a 256k-token context window, double Mistral Medium 3.1's 128k.

DeepSeek's new open flagship: about 8 times more text in memory than V3.2, while using a fraction of the computing power — and it solves about 8 in 10 real-world coding tasks.

The smaller of the two V4 preview models, with 284B total and 13B active parameters, a 1M-token context window and MIT-licensed open weights, priced at $0.14 per million input tokens and $0.28 per million output tokens.

OpenAI's GPT-5.5 roughly doubles its score on a 1M-token long-document test and solves more command-line coding tasks than GPT-5.4 — but its API price doubles.

An open 27B dense model in Alibaba's Qwen3.6 family, released six days after Qwen3.6-35B-A3B, under Apache 2.0 with a 262K native context extensible to 1M.

OpenAI's first image model that can reason before it draws: with thinking on (paid ChatGPT plans) it makes up to eight consistent images from one prompt. Up to 2K resolution in the API and much better text in Japanese, Korean, Chinese, Hindi and Bengali.

An early preview of the Max model in Alibaba's Qwen3.6 series, text-only with a 262K-token context window, offered through Qwen Studio and the Model Studio API.

xAI's update to Grok 4.20: it can now hold a whole book of text at once, watch video, and costs 40% less for the text you send in — with a few points higher on a broad skills test.

Anthropic's updated flagship: a clear jump on real-world coding and much sharper at reading charts than 4.6 — at the same price.

Alibaba's open mixture-of-experts model activates 3 billion of its 35 billion parameters and scores 51.5% on Terminal-Bench 2.0, up from 40.5% for Qwen3.5-35B-A3B; it is licensed under Apache 2.0.

New text-, image- and reference-to-video models plus a video-editing model, two days after the Wan2.7-Image model that unified image generation and editing; available through Alibaba Cloud Model Studio and the Wan website.

Google's most cost-efficient video generation model, launched in preview in the Gemini API for rapid iteration and high-volume applications.

Alibaba's first closed-weight Qwen model: a native omni-modal API model that takes text, images, audio and video, with a 256K-token context window.

Xiaomi's MiMo-V2-Omni reads text, images, video and audio in one model and can call tools and act on app screens, at $0.4 / $2 per 1M tokens through the API.

Mistral Small 4 merges Mistral's Magistral reasoning, Pixtral image and Devstral coding models into one Apache 2.0 model with 119B total and 6B active parameters, switchable reasoning effort and a 256k-token context window, double Mistral Small 3.2's 128k.

An open-weight (Apache 2.0) code agent for the Lean 4 proof assistant that generates code and formally proves implementations against specifications.

GLM-5-Turbo is a closed-weights variant of GLM-5 that Z.ai optimized for OpenClaw agent work such as tool calling, timed and persistent tasks and long-chain execution, with a 200K-token context window.

OpenAI's first general-purpose model with built-in computer use: it completes 75.0% of OSWorld-Verified desktop tasks versus 47.3% for GPT-5.2 and accepts up to 1M tokens of context, at a higher price per token.

A 15B open-weight Phi model built on Phi-4-reasoning that reads images and mixes quick answers with step-by-step reasoning, aimed at visual math and science and at finding elements on computer and phone screens.

Built for high-volume work, Gemini 3.1 Flash-Lite scores 86.9% on GPQA Diamond versus 66.7% for Gemini 2.5 Flash-Lite, but its price per 1M tokens rises from $0.10 / $0.40 to $0.25 / $1.50.

Alibaba's small open Qwen3.5 models add built-in image and video understanding, and the 9B beats the previous generation's Qwen3-30B-A3B-Thinking-2507 on 24 of 26 of Alibaba's language tests.

Brings Nano Banana Pro's quality and world knowledge to the faster Flash tier, with 512px-to-4K output and consistent identity for up to five characters — at $0.067 per 1K image, half the Pro price.

Google's updated flagship: its score on brand-new reasoning puzzles more than doubled over Gemini 3 Pro — at the same price.

xAI's new flagship Grok: the first version that retrains itself every week from real-world feedback, and in March 2026 it held the record on Artificial Analysis's test for admitting it doesn't know instead of making up an answer.

Merges Qwen-Image's separate generation and editing models into one smaller, faster model with native 2K resolution and prompts up to 1,000 tokens; available in Qwen Chat and Alibaba Cloud's API, with no downloadable weights.

Anthropic's Opus 4.6: a big jump on brand-new reasoning puzzles and the first Opus that can read 1M tokens of text at once — but its score on fixing real software bugs (SWE-bench Verified) is unchanged.

An open-weight (Apache 2.0) coding-agent model with 80B total and 3B active parameters and a 256K context, built for coding agents and local development.

xAI's image-and-video generator: 10-second, 720p clips with native audio, in the Grok app and the new Grok Imagine API.

GLM-Image replaces CogView4-6B with an open model that pairs a 9B-parameter autoregressive generator with a 7B-parameter diffusion decoder that has a glyph encoder for text inside images, and costs $0.015 per image on the Z.ai API versus $0.01 for CogView-4.

Image inputs and outputs 20% cheaper in the API than GPT Image 1.

Adds reference-to-video, so creators can appear in clips with their own look and voice, plus multi-shot videos up to 15 seconds; offered through Alibaba Cloud Model Studio and the Wan website, with no downloadable weights.

GPT-5.2 Thinking beats or ties top industry professionals in 70.9% of GDPval work-task comparisons, according to expert judges, and fixes 80.0% of SWE-bench Verified software bugs versus 76.3% for GPT-5.1, at a 40% higher API price per token.

Released as open weights under the Apache 2.0 license, with 41B active and 675B total parameters.

Ministral 3 comes in 3B, 8B and 14B sizes, each with base, instruct and reasoning versions that understand images, all under Apache 2.0, while Ministral 8B's weights were released for research use under the Mistral Research License.

Alibaba's first Z-Image model is an open-weights (Apache 2.0) image generator with 6 billion parameters that needs only 8 steps per image and fits on a 16 GB consumer graphics card.

Opus 4.5 fixes more real software bugs than Opus 4.1 (80.9% vs 74.5% on SWE-bench Verified) and costs two-thirds less, at $5 / $25 per 1M tokens instead of $15 / $75.

Built on Gemini 3 Pro: far better text rendering inside images, consistent identity across up to five subjects, and 2K/4K output.

Gemini 3 Pro fixes more real software bugs than Gemini 2.5 Pro (76.2% vs 59.6% on SWE-bench Verified) but costs more to run: $2 / $12 per 1M tokens, up from $1.25 / $10.

Microsoft's first image model built entirely in-house: it debuted in LMArena's text-to-image top 10 (Oct 2025) and launched in Bing Image Creator and Copilot.

Richer native audio, from natural conversation to synchronized sound effects, and more narrative control. Audio now also works when building a clip from reference images or start and end frames, or when extending a video, at the same API price as Veo 3.

Added synchronized dialogue and sound effects and launched with a Sora social app for putting yourself into clips, which OpenAI closed on April 26, 2026 before shutting down the Sora 2 API on September 24, 2026.

A smaller GPT Image that costs 80% less in the API.

Claude Sonnet 4.5 completes 61.4% of OSWorld computer tasks, up from 42.2% for Claude Sonnet 4, at the same $3 / $15 price per 1M tokens.

Google's image generation moves into Gemini itself: consistent characters across edits and multi-image blending, at Flash speed.

GPT-5 fixes 74.9% of real software bugs on SWE-bench Verified versus 69.1% for o3 and makes about 80% fewer factual errors on LongFact and FactScore prompts, with cheaper input but pricier output tokens.

Alibaba's open image model (Apache 2.0), strong at rendering text inside images.

Open weights again under Apache 2.0: two mixture-of-experts models (text- and image-to-video) and a 5B model that makes 720p video on a single consumer GPU. No later main version, from 2.5 on, has downloadable weights.

xAI scaled up reinforcement-learning training for Grok 4 by over an order of magnitude in compute, with native tool use and real-time search and a 256,000-token API context window, at the same $3 / $15 price as Grok 3.

Anthropic's upgrade to Claude Sonnet 3.7 for coding and reasoning, 65% less likely to take shortcuts on agentic tasks, made available to free users at the same API price.

A mobile-first open model previewed in Google AI Studio, with 5B and 8B raw sizes that run in about 2GB and 3GB of memory.

Google's video model gains synchronized audio — dialogue, ambient sound and effects generated with the footage.

Qwen3 open models can switch between step-by-step thinking and quick answers, support 119 languages and dialects, and are licensed under Apache 2.0.

A Flash model with thinking that developers can turn on or off and limit with a thinking budget, in preview through the Gemini API, Google AI Studio and Vertex AI.

o3 fixes 69.1% of real software bugs on SWE-bench Verified versus 48.9% for o1 and makes 20% fewer major errors on hard real-world tasks in external expert reviews, at a launch price a third lower.

Meta's first Llama models built as a mixture of experts and natively multimodal; Llama 4 Scout can read up to 10M tokens at once, up from 128K in Llama 3.

Gemini 2.5 Pro is a thinking model that reasons before it answers and scored 18.8% on Humanity's Last Exam without tools, but it launched with a 1M-token context window, half of Gemini 2.0 Pro's 2M.

Open-weights models in 1B, 4B, 12B and 27B sizes with a 128K-token context window, with image input on the 4B, 12B and 27B sizes.

The first Claude model that can either answer at once or think step by step first, with developers setting a thinking budget in the API, at the same price as Claude 3.5 Sonnet.

DeepSeek's reasoning model, trained from the DeepSeek-V3 base with large-scale reinforcement learning and released under the MIT license, roughly doubles V3's AIME 2024 math score at about twice V3's API price.

Alibaba's video model, first offered on its website and API on 9 Jan 2025; the weights followed on 25 Feb 2025 under Apache 2.0.

Gemini 2.0 Flash arrived as an experimental model in the Gemini API, and Google says it beats Gemini 1.5 Pro on key benchmarks at twice the speed.