2026 has produced more frontier AI models than any year before it, and most of them were out of date within a quarter. This page lists every major release in date order, with what it cost and why it mattered, so you can see the whole year at once instead of piecing it together from launch posts.
We have tracked AI launches every weekday since late March 2026 in The AI Cook daily digest, and every entry below links to the issue where we covered it in full. Dates are the day we reported each release, which is usually launch day or the morning after. Prices are list API prices per million input and output tokens at launch.
On this page
The line-up right now (October 2026)
If you only need to know what is current, start here. These are each major lab's leading models as of 5 October 2026.
| Lab | Top model | Value pick | Notes |
|---|---|---|---|
| Anthropic | Claude Fable 5.1 | Sonnet 5.5 ($2/$10) | Opus 5.5 ($4/$20) does Fable-level work for less. Mythos 5 is restricted to cyber defenders. |
| OpenAI | GPT-6 Astra | GPT-6.1 Sol ($2/$10) | GPT-5.5 retires on 14 October. GPT-6 Luna covers high-volume work. |
| Gemini 4 Argon | Gemini 3.8 Flash ($0.75/$3.75) | Argon launched for cyber defenders first, with wider access to follow. | |
| SpaceXAI | Grok 4.7 | Grok 4.7 ($2/$6) | Same price as 4.6, tuned for long coding runs. |
| Meta | Muse Spark 1.3 | Muse Glimmer (open, 30B) | Now charges for its frontier models. Muse is its consumer agent. |
| Open weights | Kimi K3, MiMo-V2.6 Pro | DeepSeek-V4.1-Flash | All MIT or similarly permissive, and all from Chinese labs. |
April 2026
The month the 2026 race properly started. Google, OpenAI, Anthropic and DeepSeek all shipped flagships within ten days of each other.
- 10 AprilGPT-5.4 opens to free usersOpenAIReasoning-class intelligence reaches the free tier of ChatGPT, setting the bar for what a free plan should include.
- 17 AprilClaude Opus 4.7AnthropicStronger coding and an image ceiling of roughly 3.75 megapixels, launched on five clouds at once. API pricing held at $5 input and $25 output per million tokens.
- 22 AprilGemini 3 Pro (preview), Gemini 3.1 Pro and Gemini 3 FlashGoogleUnveiled at Google Cloud Next. Gemini 3 Pro took first place on LMArena at 1,501 Elo and scored 91.9 per cent on GPQA Diamond.
- 22 AprilChatGPT Images 2.0OpenAIImage generation that reasons before drawing, at up to 2K resolution, rolled out to every ChatGPT user.
- 23 AprilGPT-5.5OpenAIPitched as a model for real work: plans, uses tools, checks itself and keeps going. Arrived six weeks after GPT-5.4 and reached ChatGPT, the API and GitHub Copilot within days.
- 24 AprilDeepSeek V4 (preview)DeepSeekOpen weights with a one-million-token context. V4-Pro has 1.6 trillion parameters (49 billion active), V4-Flash 284 billion (13 billion active).
May 2026
A month of agents rather than models. The big labs spent it turning chatbots into things that act, and Anthropic closed it with another Opus.
- 6 MayGPT-5.5 Instant becomes ChatGPT's defaultOpenAIThe fast variant quietly replaced the old default for everyday chats.
- 15 MayGrok Build coding agentxAIxAI's answer to Claude Code and Codex: a terminal agent running up to eight workers, launched behind a $299 SuperGrok Heavy plan.
- 20 MayGemini Spark and Gemini OmniGoogleGoogle I/O reframed Gemini as an always-on agent that can carry out tasks inside other apps, not just answer questions.
- 28 MayClaude Opus 4.8AnthropicSWE-bench Pro rose from 64.3 to 69.2 per cent, and a Fast mode ran about 2.5 times quicker, with no price increase.
- 29 MayGrok Build 0.1 on the APIxAIA 100-tokens-per-second coding model at $1 input and $2 output per million tokens, well under the frontier price.
June 2026
June brought the year's biggest single jump: Anthropic made a Mythos-class model public. It also showed how short a model's life had become.
- 3 JuneMAI-Thinking-1MicrosoftMicrosoft's first in-house reasoning model, unveiled at Build alongside MAI-Image-2.5. A clear signal it wants options beyond OpenAI.
- 8 JuneGemma 4 QAT checkpointsGoogle DeepMindQuantisation-aware versions of the open Gemma 4 family. The smallest text model now runs in under 1GB of memory.
- 10 JuneClaude Fable 5 and Claude Mythos 5AnthropicFable 5 became the first Mythos-class model open to the public, at $10 input and $50 output per million tokens. The unrestricted Mythos 5 went only to vetted cyber defenders.
- 16 JuneClaude Sonnet 4 and Opus 4 retiredAnthropicRetirements arrived as fast as launches. OpenAI had retired GPT-5.2 days earlier.
July 2026
The busiest month of the year for flagships, and the month price became the headline. OpenAI and xAI launched hours apart, Meta started charging, and Opus 5 halved the cost of near-top intelligence.
- 1 JulyClaude Sonnet 5AnthropicBecame the default for Free and Pro users, closing much of the gap to Opus 4.8 on agentic work at a fraction of the cost.
- 9 JulyGPT-5.6 Sol, Terra and LunaOpenAIThree tiers, three prices: Sol $5/$30, Terra $2.50/$15 and Luna $1/$6 per million tokens. Launched with GPT-Live, OpenAI's first full-duplex voice model.
- 9 JulyGrok 4.5xAIWent public within hours of GPT-5.6, with big claims but no system card or benchmark table.
- 9 JulyMuse Spark 1.1MetaMeta's first paid model ($1.25/$4.25 per million tokens), and a clean break from its open-weights-only history.
- 22 JulyGemini 3.6 Flash and Flash CyberGoogleA cheaper Flash, plus a security-tuned twin. Gated cyber models became a pattern from here on.
- Late JulyClaude Opus 5AnthropicClose to Fable 5 intelligence at roughly half the price, holding Opus pricing at $5/$25. Became the default on Claude Max.
- Late JulyKimi K3Moonshot AIThe largest open-weight model yet at 2.8 trillion parameters, with only 16 of its 896 experts active per token. Briefly topped the coding arena.
August 2026
No new top-tier flagship from the big three, so the action moved to price cuts and open weights, most of them from China.
- 3 AugustDeepSeek V4 Flash 0731DeepSeekMIT-licensed open weights that went straight into the top three open models.
- 4 AugustQwen3.8-MaxAlibabaAlibaba's flagship, released alongside MiniMax's open-source H3 video model as Asian labs set the open-weight pace.
- 11 AugustGPT-5.6-Cyber and Muse GlimmerOpenAI, MetaOpenAI gated its cyber model behind vetted tiers. Meta open-sourced Muse Glimmer, a 30B agentic model that runs on a single GPU.
- 13 AugustGrok 4.6SpaceXAIFrontier-level index scores at $2 input and $6 output per million tokens, about half the price of comparable rivals.
- 14 AugustGemini 3.7 FlashGoogleHalf the price of 3.6 Flash at $0.75/$3.75 per million tokens. DeepSeek V4-Pro went generally available the same day.
- 17 AugustQwen3.8-27BAlibabaA dense multimodal model under Apache 2.0 that can run locally and be used commercially without caveats.
- 27 AugustQwen3.8-Flash-Next and GLM-5.3-FlashAlibaba, Z.aiA preview of the Qwen 4 architecture, and an MIT-licensed GLM that undercut on price.
September 2026
The densest month of 2026. OpenAI moved to GPT-6, Anthropic refreshed all three tiers, and cheaper challengers from DeepSeek, StepFun and Xiaomi kept closing the gap.
- 2 SeptemberClaude Fable 5.1AnthropicCache reads 75 per cent cheaper, which works out at about 25 per cent off a typical workload. Qwen's 2.4 trillion parameter Max-0902 undercut it within hours.
- 3 SeptemberGemini 3.8 Flash and Muse Spark 1.3Google, MetaGoogle kept its monthly Flash cadence and again locked the cyber variant behind vetting.
- 4 SeptemberGPT-6 AstraOpenAIBuilt around computer use. OpenAI claimed 98 per cent on FrontierMath Tier 4 and 99.9 per cent on ARC-AGI-3, and talked openly about AGI. Reached every paid tier and the API by 7 September.
- 9 SeptemberMuseMetaA personal agent running on Muse Spark 1.3 that books, buys and fills in forms inside its own sandboxed virtual machine.
- 11 SeptemberDeepSeek-V4.1-Flash and SWE-2DeepSeek, CognitionDeepSeek's MIT-licensed 552B model (8B active) beat its own V4-Pro, which was retired. Cognition's SWE-2 came within a point of frontier coding for 64 per cent less.
- 21 SeptemberStep 5 PreviewStepFunA 600B mixture-of-experts model with a 1M context at $1/$2.70 per million tokens, roughly a seventh of GPT-5.6 Sol's price.
- 22 SeptemberMiMo-V2.6 Pro and Grok 4.7Xiaomi, SpaceXAIXiaomi's MIT-licensed 1.02 trillion parameter model claimed the top open-weight spot. Grok 4.7 was a same-price upgrade tuned for hours-long coding sessions.
- 23 SeptemberClaude Opus 5.5, GPT-6 Sol and GPT-6 LunaAnthropic, OpenAISame afternoon, same pitch. Opus 5.5 does Fable-level work for 40 per cent less than Opus 5 ($4/$20). GPT-6 Sol and Luna halved the API price of the models they replaced.
- 29 SeptemberClaude Sonnet 5.5AnthropicNear-Opus coding on the everyday tier, more than 30 per cent faster than Sonnet 5, at an unchanged $2/$10.
- 30 SeptemberGPT-6.1 SolOpenAIReplaced GPT-6 Sol after a week. Matches Astra on OpenAI's coding benchmark at a fifth of the price ($2/$10).
October 2026
So far: Google's long-awaited Gemini 4, and a serious open model from Europe.
- 1 OctoberGemini 4 ArgonGoogle DeepMindGoogle's most powerful model, with a one-million-token output limit and 77.9 per cent on DeepSWE. Released first to cyber defenders only.
- 5 OctoberKolibriAleph AlphaAn Apache 2.0 German-English model (78B, about 3.5B active) built for data that has to stay on your own servers.
Five patterns from the year so far
1. Flagships now last weeks, not years
GPT-5.5 arrived six weeks after GPT-5.4 and is being retired on 14 October, less than six months after launch. GPT-6 Sol was replaced in seven days. Anthropic shipped six top-tier models (Opus 4.7, Opus 4.8, Fable 5, Opus 5, Fable 5.1, Opus 5.5) in six months. If you build on a specific model name, plan for it to change at least once a quarter.
2. The launch headline is now the price
By September, releases led with the invoice rather than the benchmark. Opus 5.5 dropped Opus pricing from $5/$25 to $4/$20. GPT-6.1 Sol offers near-Astra performance at a fifth of the cost. Gemini 3.7 Flash halved the price of its predecessor. Caching discounts, which most people never look at, delivered some of the biggest real-world savings of the year.
3. The most capable models are increasingly gated
Claude Mythos 5, GPT-5.6-Cyber, the Gemini Flash Cyber models and Gemini 4 Argon all went to vetted security teams before (or instead of) the public. Expect the strongest version of each new generation to reach defenders first.
4. Open weights went east
Almost every leading open-weight release came from a Chinese lab: DeepSeek, Moonshot, Alibaba, Xiaomi and Z.ai. Meta moved the other way and started charging. Europe's notable entries were Mistral and, this month, Aleph Alpha's Kolibri. For a business that needs to run a model on its own servers, the choice is better than ever, but licences vary, so read them.
5. Coding is the benchmark that sells
Nearly every launch led with a coding or agentic score: SWE-bench Pro, Terminal-Bench, DeepSWE, FrontierCode. That is where the paying customers are. If you want to see which tools turn those models into something useful, read our guide to the best AI coding tools in 2026.
Keeping up without the noise: this timeline is updated as new models ship. For the day-by-day version, with what each release means for your work, subscribe to the daily digest.
Frequently asked questions
What is the most powerful AI model right now?
As of October 2026, the top generally available models are OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1. Google's Gemini 4 Argon leads several benchmarks but launched to vetted cyber defenders first. Which is best depends on the task: coding, long agentic work and document-heavy jobs each have different leaders.
Which AI model offers the best value?
The best price-for-capability models in October 2026 are Claude Sonnet 5.5 and GPT-6.1 Sol (both $2 input and $10 output per million tokens), Grok 4.7 ($2/$6), and Gemini 3.8 Flash ($0.75/$3.75) for high-volume work. Open-weight models such as DeepSeek-V4.1-Flash are cheaper still if you can host them.
What is the best open-source AI model in 2026?
Xiaomi's MiMo-V2.6 Pro and Moonshot's Kimi K3 are the strongest open-weight models as of October 2026, with DeepSeek-V4.1-Flash the most efficient to run. Qwen3.8-27B (Apache 2.0) is the best fit for running locally on a single machine.
How often are new AI models released?
Very often. In 2026 the major labs have shipped a significant model roughly every six weeks each, and across the industry a notable release lands almost every week. September 2026 alone saw new flagships or major updates from Anthropic, OpenAI, Google, Meta, DeepSeek, Xiaomi and SpaceXAI.
Why do AI companies retire old models so quickly?
Serving a model ties up expensive GPU capacity, so labs move users to newer models that are usually cheaper to run as well as more capable. In 2026 most models have been retired or replaced within six to twelve months of launch, and some within weeks.
What to take from this
The model you picked in spring is probably not the best choice now, and the one you pick today will not be in six months. That is less of a problem than it sounds: prices are falling faster than capability is rising, so most upgrades pay for themselves. Pick a model for the job you have, keep your prompts and tests portable, and check back here when the next one lands.
New models every week. One email that tells you which ones matter.
We read every launch so you do not have to, and send the short version every weekday morning before 8am. Free, no hype, unsubscribe any time.
You're in. First issue lands tomorrow morning.