Get occasional email updates on new AI tools. No spam.
Each lab's newest release, with the facts that matter. Release date, open or closed weights, context window, API price and what it accepts. No scores, no rankings. This is a living snapshot, not an archive: when a lab ships a successor, the older entry comes off.
This is not a leaderboard. We list each lab's most recent models and the facts printed on the vendor's own pages, newest first. Prices are list prices in USD per million tokens (input / output) and change often. “Not confirmed” means we couldn't read it on an official page, so we left it blank instead of guessing. “Limited access” and “Preview” mean you may not be able to use the model yet, even if a price is published.
Rolling out first to trusted cyber defenders in Google's Fairwind Program, with paid API customers and Google AI Ultra subscribers next. It is not yet in the public Gemini API model list.
A cheaper tier released after Astra. OpenAI positions it as near-Astra capability at lower cost.
Anthropic's Opus-tier model, succeeding Claude Opus 5.
Built on a new, larger base model than Grok 4.6, at the same price and speed.
DeepSeek says it replaces V4-Pro; V4-Pro requests are routed to V4.1-Flash at Flash rates from Sep 14.
OpenAI's most capable model. It launched to trusted organizations first, with ChatGPT paid plans, the API, Azure and Amazon Bedrock following within days, per OpenAI.
A generally available Gemini model, succeeding Gemini 3.7 Flash.
Meta's frontier model is closed-weights, available through the Meta Model API. It succeeds Muse Spark 1.2.
Anthropic's current flagship, succeeding Claude Fable 5. A version of the same model with fewer safeguards, Claude Mythos 5.1, is limited to vetted organizations.
IBM's open Granite family, succeeding Granite 4.1.
Same base as GLM-5.2 with post-training gains. GLM-5.2 was MIT-licensed; 5.3 moved to a custom license.
Meta's newest open-weight model. The newest Llama we could verify is Llama 4, from April 2025.
Alibaba's largest model to date. Its base model, Qwen3.8-2.4T-A95B, is published on Hugging Face.
Succeeds Kimi K2.6 and K2.7; Moonshot's docs recommend upgrading to K3. The license is a custom one, not MIT.
Announced at Volcano Engine's Force conference; a preview post appeared on ByteDance Seed's site on Jun 19.
NVIDIA's Ultra-tier Nemotron 3 model. A smaller Nemotron 3.5 Lightning (30B) followed on Aug 11.
Microsoft's first in-house reasoning model, announced at Build and offered in preview on Azure.
Succeeds MiniMax M2.7. Weights were promised within about 10 days of launch.
A 218B mixture-of-experts model (25B active) that consolidates Command A Reasoning, Vision and Translate.
Mistral's docs call it its newest flagship.
Not listed yet. Amazon (Nova) — No current flagship could be confirmed from AWS's own pages; its newest verifiable models date from December 2025. We would rather leave a lab off than list a model we can't back up.
A short list of rules, so you know what a blank or a date means.
Vendor pages first — every card links the lab's own page for that model. Specs and prices come from those pages, not from leaderboards or reposts.
Blank beats a guess — if a lab's pages don't state a context window or price, the field says “Not confirmed.” A “Partly confirmed” tag means a date or figure rests on a secondary source or conflicting pages; the note under the card says which.
Latest only — each lab shows its newest release. When a successor ships, the older entry is replaced, so this page stays readable as the pace of releases keeps climbing.
No scores, no ranking — we don't publish benchmark numbers or order models by quality. The only ordering here is release date, newest first.