Tracker
AI model releases 2026
Every AI model launch, preview and reported plan from our reporting, newest first, with prices and the sourced story behind each entry. Updated with every edition.
Prices and claims are what each story reports; a company’s own benchmark stays a company claim. The claim check compares each maker’s main claim with an independent test that our story cites. Released models so far: 13 backed, 13 mixed, 15 not checked yet.
Filter the list
November 2026
The fast, everyday model of the Gemini 4 family, after Gemini 3.8 Flash. In the Gemini app, Search, AI Studio and the Gemini API, and in Antigravity for paying subscribers.
Claim check: BackedArtificial Analysis scores it 51 at the default setting and 54 at the highest, against 50 for Gemini 4 Argon. Google’s speed and accuracy gains over Flash are its own.
Price $1.50 / $7.50 per million input / output tokens
Google’s Gemini 4 Hydrogen beats Argon on an outside testA 27-billion-parameter world action model that turns camera images and text into robot movements. Open weights and code under Apache 2.0; first shown to partners in September.
Claim check: Not checked yet
Runway opens its Praxis-1 robot model to everyoneNow open to paying developers in the Gemini API, to businesses, and to Google AI Ultra subscribers in the Gemini app, AI Mode and Antigravity, with a /swarm agent-team preview.
Claim check: BackedArtificial Analysis scores it 60, fourth of 74 models, behind Claude Fable 5.2 at 65. Google’s other test rankings are its own.
Price $2 / $10 per million input / output tokens to Nov 30, then $4 / $20
Google opens Gemini 4 Carbon to paying usersAnthropic’s top model for all users, with safeguards for cybersecurity and biology. Anthropic says it writes answers up to 30% faster than Fable 5.1.
Claim check: BackedArtificial Analysis scores it 65, among the leading models, but calls it particularly expensive. The Fermat and physics results are Anthropic’s own.
Price $8 / $40 per million input / output tokens, 20% below Fable 5.1
Anthropic cuts Fable’s price 20% with its new 5.2 modelsThe same model as Fable 5.2 with fewer cybersecurity and biology safeguards, only for vetted organisations through Anthropic’s verification programs and Claude Security.
Price $8 / $40 per million input / output tokens
Anthropic cuts Fable’s price 20% with its new 5.2 modelsMeta’s smallest Muse model, built for the work AI agents repeat all day. On the Meta Model API and in Muse Code; its weights are not public.
Claim check: BackedArtificial Analysis scores it 36, level with Claude Haiku 5.5 (max), at about a third of the cost per task. It is weaker on facts.
Price $0.10 / $0.50 per million input / output tokens
Meta’s Muse Ember matches Haiku 5.5 at a third of the costA 594-billion-parameter model, with 17 billion active, on the company’s Tinker service. The company plans to release the weights under Apache 2.0 after outside safety testing, so they are not public yet.
Claim check: BackedArtificial Analysis scores it 31, the same score the company gives, and level with Gemini 3.8 Flash (high). The “open” part is still a plan.
Price $0.30 / $1.00 per million input / output tokens
Thinking Machines launches Inkling 2 Small, weights to comeA security version of GPT-6 Astra, for a select group of approved Daybreak Red customers through the API and Codex. OpenAI rates it Critical for cybersecurity.
Claim check: MixedOpenAI says it beats GPT-6 Astra on most cyber tasks. Its system card says outside tester Irregular found the two broadly comparable, with better results on long attack tasks.
OpenAI limits GPT-6 Cyber, its new bug-hunting modelTencent’s new model adds image reading and computer use, in its own apps and API and on outside platforms. Tencent says the weights will follow “in the coming days”; only the August preview can be downloaded so far.
Claim check: Not checked yet
Price About $0.90 / $2.70 per million input / output tokens, Tencent says
Kimi K3.1 Pro’s weights are out, and Tencent launches Hy4Moonshot published the full weights on Hugging Face, as it promised, under its own licence. Businesses selling model access with over $20 million in yearly revenue need a separate deal.
Claim check: MixedArtificial Analysis puts it first on Terminal-Bench 4.0, at 67%, which backs Moonshot’s coding claim. But it scores 46 overall on Index v5, the highest for an open model and 12 points behind Claude Opus 5.5.
Kimi K3.1 Pro’s weights are out, and Tencent launches Hy4
October 2026
Mistral’s 1-trillion-parameter model, with 49 billion parameters active per token, is now free to download under Mistral’s own modified MIT licence. Firms that sell it as a service need a separate licence.
Claim check: MixedArtificial Analysis scores it 26 overall and 48 on coding agents, one point behind GLM-5.3, as VentureBeat reports; Mistral calls the coding result a tie.
Price $0.70 / $2.10 per million input / output tokens; weights free to download
Mistral releases the weights of its Large 4 modelAn open-weight model that transcribes recordings in 40 languages, translates 31 to or from English by Nvidia’s count, and labels up to eight speakers. Free for commercial use under the OpenMDW-1.1 licence.
Claim check: Not checked yet
Price Free to download
Nvidia’s open Nemotron model transcribes 40 languagesGoogle’s strongest model, built for cybersecurity and long tasks. Given first to some cyber defenders, then to more trusted partners; no public app or API yet.
Claim check: BackedArtificial Analysis scores it 59 on its Intelligence Index v5, first of 52 models, one point above Claude Opus 5.5.
Price $2 / $10 per million input / output tokens at launch; later $4 / $20
Google’s Gemini 4 Carbon tops a new independent testA world model that draws interactive scenes in real time, with sound and talking characters, steered by text prompts. Research preview for trusted testers; Google AI Ultra subscribers and Gemini API partners are due “soon”.
Price No price yet
Google’s Genie 4 builds worlds with sound you can steerA fast, cheaper successor to GPT-5.6 Terra for computer use, visual reasoning and writing, in the API and in ChatGPT Work and Codex. OpenAI says it matches GPT-6 Astra on office tasks at one-tenth of the token price.
Claim check: BackedArtificial Analysis scores it 48, 25th of 221 models, 6 points above GPT-5.6 Terra at $0.40 per task instead of $1.40.
Price $1 / $5 per million input / output tokens; cached input $0.05
OpenAI’s GPT-6.1 Terra costs less and scores higherMoonshot’s new flagship, with a 1-million-token context and a faster Highspeed version, in Kimi’s apps and API. Moonshot says it nearly matches Claude Opus 5.5 on a coding test and promises the weights by November 4.
Claim check: MixedArtificial Analysis scores it 55, fourth of 215 models, but found it slower than average and very wordy, and lists it as proprietary until the weights are out.
Price $2 / $10 per million input / output tokens
Moonshot launches Kimi K3.1 Pro, open weights due Nov 4A “code execution model” that rewrites a program’s machine code while it runs, for CPUs and GPUs. doubleAI says pipelines ran more than 38 times faster; no outside test yet.
Price Early access on request; no public price
doubleAI says its Maxwell model makes code run fasterAn agent model with up to 1 million tokens of context, for text and images. MiniMax says it is nearly three times faster and 70% cheaper to use than before. Weights are due later.
Claim check: MixedArtificial Analysis scores it 42, fourth of 173 models, but measured 120.1 tokens per second, not 200+, and a 93% hallucination rate.
Price $0.15 / $0.60 per million input / output tokens
MiniMax launches M3.1 Flash at half the price of M3A 27-billion-parameter model on Google’s Gemma 3, made with AI Singapore, for Indonesian, English, Javanese, Sundanese, Balinese and Batak Toba. GoTo says it scored highest on regional languages in its own tests, with no scores published.
Claim check: Not checked yet
Price Free weights under the Gemma licence
GoTo releases a free AI model for Indonesian and JavaneseA “minor update” to V4.1 Flash with open weights under the MIT licence and up to 1 million tokens of context. DeepSeek says it scores higher on coding and agent tests, works faster and writes shorter answers.
Claim check: Not checked yet
Price $0.30 / $1.20 per million input / output tokens (peak, unchanged)
DeepSeek releases an open update of its V4.1 Flash model Try it yourself: who can use it, the price, an example promptPre-training finished: 2.8 trillion parameters, 80 billion active per token. Now in post-training, with no release date. Microsoft says it beats every open-weight model it tested and plans to license the weights to trusted partners.
Microsoft finishes pre-training its 2.8T MAI-Base-2 modelMeta’s open-weight model: 816 billion parameters, about 42 billion active per token, and up to 1 million tokens of context. Meta calls it the most capable US open-weight model. Large API sellers need a separate agreement.
Claim check: MixedArtificial Analysis gives it 39 points; Meta’s own table puts two Chinese open models ahead on coding tests.
Meta releases Muse Spark Open, an 816B open-weight model Try it yourself: who can use it, the price, an example promptNow open to AI Pro and Ultra subscribers in the Gemini app and in Search, and to developers on the paid Gemini API. Free users and AI Plus don’t get it yet.
Claim check: MixedArtificial Analysis scores it 53, eighth of 216 models, with Claude Opus 5.5 ahead. Google’s claim of clearer answers is untested.
Price $2 / $10 per million input / output tokens until November 30, then $4 / $20
Google opens Gemini 4 Argon to paying users and developers Try it yourself: who can use it, the price, an example promptMeta’s first frontier models, built for agents on long tasks. Muse Wonder is in the Meta Model API and Muse Code; Muse Wonder Full only in Muse Code’s paid plans. Meta says Wonder will soon power its Muse app.
Claim check: MixedArtificial Analysis scores Wonder Full 56, two points behind Claude Opus 5.5, but weaker on facts. Meta’s “state-of-the-art” agent results are its own tests.
Price $4 / $20 per million input / output tokens (Muse Wonder)
Meta’s Muse Wonder models close in on the AI frontier Try it yourself: who can use it, the price, an example promptA small decision model built on Alibaba’s Qwen3.5-9B that scores a fixed list of options, as a rival to Jev. Microsoft says it is the most accurate on 36 tests; no outside test yet.
Claim check: Not checked yet
Price $0.042 per million input tokens; output is free
Microsoft launches Decision-1, a fast model for AI choicesAnthropic’s new small model, the first Haiku with effort controls. In the Claude apps, Claude Code, the API and the big clouds.
Claim check: BackedArtificial Analysis scores it 43 on its Intelligence Index, second of 182 models in its class, though it uses many tokens. Anthropic’s benchmark gains are its own.
Price $0.10 / $0.50 per million input / output tokens up to 100K tokens; $0.50 / $2.50 above
Anthropic’s Haiku 5.5 cuts small-model prices by 90% Try it yourself: who can use it, the price, an example promptNow in ChatGPT’s Chat tab with Intelligent UI: Sol for Plus, Pro, Business and Enterprise from October 7, Luna for Free and Go from October 8.
ChatGPT gets GPT-6 and answers with charts and buttons Try it yourself: who can use it, the price, an example promptA model with 1 trillion parameters, 49 billion used per token, reading text and images. Public preview through Mistral’s API; Mistral says the weights follow by the end of October.
Claim check: MixedArtificial Analysis scores the preview 38 on its Intelligence Index, 64th of 225 models: above the middle score of 26, but far from the leaders. Mistral’s cyber ranking is its own claim.
Price $1.36 / $4.18 per million input / output tokens (preview)
Mistral previews Large 4, a 1-trillion-parameter modelA text model with 501 billion parameters, 23 billion used per token, and up to 1 million tokens of context. Reflection says it matches GLM-5.2 on hard reasoning with 3–4× less computing; no outside test yet.
Price Free weights (Apache 2.0) due later in October; early-access list for now
Reflection unveils Beam, an open model to rival China’s labsMicrosoft’s first live speech-to-text model, in 60 languages, in public preview in Microsoft Foundry. Two text-to-speech models, MAI-Voice-2.1 and 2.1-Flash, came with it.
Claim check: BackedOn Artificial Analysis’s streaming leaderboard it gets 2.5% of words wrong at the first partial transcript, the lowest of 32 models shown; the next is 3.4%.
Price $0.54 per audio hour (introductory, until the end of 2026)
Microsoft’s live transcription model tops an outside testTwo decision models for AI agents (27B and 9B) that also read images, offered as an open rival to Jev. Cloudflare says a Clef model tops 7 of 10 decision tests; no outside test yet.
Claim check: Not checked yet
Price $0.24 / $0.09 per million input tokens on Workers AI; free weights (Apache 2.0)
Cloudflare releases open Clef models to take on JevA small decision model that picks one option from a fixed list for AI agents, as an open alternative to Jev. AWS says it takes about 115 ms on one graphics card; no outside test yet.
Claim check: Not checked yet
Price Free (Apache 2.0)
Amazon releases a free, open rival to the Jev decision model
September 2026
Google’s new flagship, for now only for trusted cyber defenders in its Fairwind Program. Google says it leads on 12 of 18 tests and can write up to 1 million tokens at once. No date for a wider release.
Claim check: MixedArtificial Analysis scores it 53, level with GPT-6 Astra and five points behind Claude Opus 5.5, at $1.99 per task.
Price $2 / $10 per million input / output tokens at first, later $4 / $20
Google’s Gemini 4 Argon goes to cyber defenders firstAn upgrade to GPT-6 Sol. OpenAI says it nearly matches GPT-6 Astra on coding, computer use and office work at a fifth of Astra’s token prices.
Claim check: BackedArtificial Analysis scores it 52, against 53 for GPT-6 Astra, at $0.72 per task instead of $3.26.
Price $2 / $10 per million input / output tokens; cached input $0.10
Independent test puts GPT-6.1 Sol a point behind AstraThe second Claude 5.5 model. Anthropic says it is more than 30% faster than Sonnet 5 and costs up to 30% less per task.
Claim check: MixedArtificial Analysis puts it two points behind Opus 5.5. At maximum effort it lists $5.46 per task (October 8), about 7% more than Sonnet 5; Anthropic’s savings figures refer mostly to lower effort settings.
Price $2 / $10 per million input / output tokens
Anthropic’s Sonnet 5.5 nears Opus at half the token pricePlanned for release within weeks; cancelled after internal tests, because it fell short on staying within scope and on how it reports its work.
OpenAI scrapped GPT-6.1 Astra after it fell short on safetyHas entered post-training; DeepMind’s new chief wants an early version out “much earlier” than the end of the year. No date given.
Google’s Gemini 4 has entered post-trainingA 7-billion-parameter model that turns camera images and text into robot movements. Black Forest Labs says it tops the RoboLab-120 benchmark. Weights free for non-commercial use.
Claim check: BackedIt ranks first on NVIDIA Research’s public RoboLab-120 leaderboard, at 42.9%. We could not confirm whether NVIDIA re-ran the result itself.
Image AI maker Black Forest Labs released a robot modelVoice agents with a lip-synced video face for Gemini Enterprise customers, in 97 languages, with a SynthID watermark on every video.
Google’s Gemini can now talk to customers as a video avatarA fourth security-focused model this year, in testing with a few Daybreak Red customers, according to Fortune. OpenAI has not confirmed it.
OpenAI is preparing GPT-6 Cyber, Fortune reportsText-to-speech in more than 100 languages; can clone a voice from a short sample after the owner records a consent statement.
Claim check: MixedArtificial Analysis ranks Flash second in its voice arena. The first places Google cites come from tests by Hume AI, whose founder co-wrote Google’s post.
Price About $0.54 to $0.81 per hour of audio until the end of the year
Google’s new Gemini voices can copy a person’s voiceTwo cheaper tiers below GPT-6 Astra in the API. OpenAI’s own tables put Sol ahead of Claude Opus 5 on AutomationBench.
Claim check: MixedArtificial Analysis finds them about level with GPT-5.6, at roughly half the cost per task: cheaper, not smarter.
Price $2 (Sol) and $0.10 (Luna) per million input tokens, half of GPT-5.6’s prices
OpenAI’s GPT-6 Sol and Luna cut API prices in halfAnthropic says it performs like its more expensive Fable 5.1 on most work.
Claim check: BackedArtificial Analysis ranked it first of the 206 models it tracked at launch, with 58 points. Most other figures are Anthropic’s.
Price $4 per million input tokens, $20 per million output tokens, 20% below the previous price
Claude Opus 5.5 cuts token prices by 20%OpenAI says an internal model in training since August 28 has solved more than 100 open maths problems. None are named or published.
OpenAI says a new model solved over 100 open maths problemsA reasoning model for CRM work, based on Nvidia Nemotron and trained on synthetic business tasks. Its strongest results come from Salesforce’s own tests.
Claim check: Not checked yet
Salesforce built a reasoning model for CRM workA small model that publishes checkpoints, data, code and hashes for every training step, so its training can be replayed. About five times slower to run.
Claim check: Not checked yet
Gensyn made a small model whose training can be replayedA model that returns choices and probabilities instead of text, from former OpenAI researchers, with $40 million in seed funding.
Claim check: Not checked yet
Price $0.042 per million input tokens; output is free
Former OpenAI researchers launched Jev, a model that decidesThe English beta started on September 14: Siri can use messages, email, photos, screen content and app actions, with device, region and daily cloud limits.
Siri AI is finally on real devicesKeeps important details during repeated edits and cuts waiting time; new sketch and comment tools guide the result.
Claim check: Not checked yet
ChatGPT Images 2.5 makes editing fasterA free research map of every possible one-letter change in the human genome, about 9 billion of them. A prediction is not a diagnosis.
Claim check: Not checked yet
AlphaGenome Atlas maps 9 billion DNA changesAvailable through Amazon Bedrock APIs with a large context window; buyers should review retention and permission settings.
GPT-6 Astra is now on Amazon BedrockA 2.5-billion-parameter long-context model for local and on-device use; much of its training data is shared. OpenBMB says it leads open models of its size.
Claim check: BackedArtificial Analysis gives it 15 points, the highest among open-weight models under four billion parameters.
MiniCPM5-2B packs long-context AI into a much smaller modelBetter computer-use skills, and OpenAI’s first widely released model with a “Critical” cyber rating.
Claim check: Not checked yet
GPT-6 Astra is powerful enough to need stricter cyber rulesMeta says the agent uses fewer tool calls and fewer tokens than the earlier version.
Claim check: Not checked yet
Meta says its new AI agent can finish work with fewer steps
How this tracker works
An entry appears when one of our stories covers a model launch, a preview, a reported plan, or a model becoming available on a new platform. A claim is Backed when an outside test supports it, Mixed when it holds only in part or only in some settings, and Disputed when an outside test contradicts it. The date is the one the story gives; when a story gives none, it is the story’s own date. We add entries with each edition and correct them under the story when a fact changes.
Reusing this list: you may quote it with a link to this page. Facts belong to the sources each story cites.