AI Model Release Tracker: New LLMs, Dates, Licences and Context

This page tracks notable large language model releases from July 2026 onwards: the date, the vendor, the context window, and either the access terms (for hosted models) or the licence (for open weights). It is for engineers who need to answer "what shipped, when, and can I use it" without reading six changelogs. Every row links to the vendor page or model card I checked. Last checked: 12 November 2026.

Prices are deliberately not here. They live on the LLM API pricing tracker.

Latest changes

• 12 Nov 2026: Google moved Gemini 4 Hydrogen (gemini-4-hydrogen) to general availability on the Gemini API and made it the default model for its Antigravity managed agent (source).
• 10 Nov 2026: Gemini 4 Carbon became generally available, ending the defenders-only period that began on 28 October (source). My Gemini 4 post covers the prices.
• 9 Nov 2026: Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, the same underlying model with different safeguards (source). Details in my Fable 5.2 write-up.
• 5 Nov 2026: OpenAI added GPT-6-Cyber, restricted to organisations with Daybreak Red approval (source). Mistral's docs give the same day as the release date for Mistral Large 4 as a generally available API model (source).
• 2 Nov 2026: Moonshot AI's Kimi K3.1 Pro weights repository went up on Hugging Face (source). Covered in my Kimi K3.1 Pro post.
• 30 Oct 2026: OpenAI listed GPT-6 Rosalind, a life sciences model limited to its trusted-access programme (source).
• 29 Oct 2026: Mistral published the Mistral Large 4 weights. The first commit in the Hugging Face repository, titled "Initial release", carries that date (source). Covered in my Mistral Large 4 post.
• 26 Oct 2026: OpenAI released GPT-6.1 Terra (source). Covered in my GPT-6.1 Terra post.
• 17 Oct 2026: DeepSeek moved its deepseek-flash API alias to the V4.1 Flash 1017 build and published matching MIT weights (source, weights).
• 15 Oct 2026: Meta released Muse Spark Open, the open-weight version of Muse Spark, under its own Muse Model License (source).

Hosted and API model releases since July 2026

These are models you reach through a vendor API. "Released" is the date in the vendor's changelog or announcement. Where a model had a restricted preview first, the date is general availability and the preview is noted under access. A model that ships both as an API and as weights is listed once, in the open-weight table below.

Two things stand out when the rows sit side by side. A context window of about one million tokens is now the default at Anthropic, Google and OpenAI, with xAI (its release notes credit the Grok models to SpaceXSI) at 500k. And three vendors have put a cyber-capable model behind a vetting programme: Mythos at Anthropic, GPT-6-Cyber at OpenAI, and Carbon at Google for the 13 days before it went GA.

Open-weight model releases since July 2026

These are models whose weights you can download. Dates are the vendor's stated release date where the model card, changelog or announcement gives one. Rows marked (HF) use the creation date of the Hugging Face repository instead, which can differ from the public announcement by a few days. Context windows are the figures on the model card. Read the licence file before you build on any of the custom licences: several restrict Model-as-a-Service use.

The licence column is the one to read slowly. Of the seventeen rows, eight are plain MIT or Apache 2.0 (counting GLM-5.3 Flash, which shares a row with its custom-licensed sibling). The largest models from Moonshot, Mistral, Z.ai and Alibaba come with custom terms that start from MIT wording and then add a clause about hosting the model for others. That clause is irrelevant if you run a model inside your own product, and decisive if you plan to resell inference. Meta's release is the only one here that is gated: you fill in a form before Hugging Face lets you download it.

A context figure on a model card is also a ceiling, not a promise. Several cards say the usable window depends on GPU memory and the serving configuration, and Qwen documents a 262,144-token native window that reaches a million only with RoPE scaling.

How this page is maintained

I update this page when a vendor changelog, release post or model card shows a new general-purpose language model, and I recheck the existing rows at the same time. Sources are primary only: vendor documentation, vendor announcements and the vendor's own Hugging Face repository, never a news report or a leaderboard. Prices, benchmark scores, image, audio, video, embedding and time-series models, community quantisations and fine-tunes are deliberately excluded. A row that I cannot confirm on a page I fetched myself is left out, not guessed. Corrections are welcome at hello@ajayk.sh.

FAQ

What is the newest AI model released in November 2026?

As of 12 November 2026, the most recent hosted release is Google's Gemini 4 Hydrogen, which became generally available that day, two days after Gemini 4 Carbon and three days after Anthropic's Claude Fable 5.2. On the open-weight side the newest notable repository is Moonshot AI's Kimi K3.1 Pro, created on Hugging Face on 2 November. The Latest changes list at the top of this page is where new entries appear first.

Which open-weight LLMs released in 2026 use an MIT or Apache 2.0 licence?

Among the releases tracked here since July 2026, the DeepSeek V4 Pro and V4.1 Flash weights and GLM-5.3 Flash are MIT, and Tencent's Hy4 preview, IBM's Granite 4.2, Qwen3.8-27B and LG's K-EXAONE 2.0 are Apache 2.0. Kimi K3.1 Pro, Mistral Large 4, GLM-5.3, the larger Qwen3.8 models, Muse Spark Open, ERNIE 5.5 Flash Base, Step 5 Preview and Solar Open 2 use custom licences. NVIDIA's Nemotron uses OpenMDW-1.1, a standard model licence that is neither MIT nor Apache.

What context window do the latest LLMs have?

Most current flagship models document a context window of roughly one million tokens: 1M at Anthropic, 1,048,576 at Google and 1,050,000 at OpenAI. xAI's Grok 4.5, 4.6 and 4.7 are listed at 500k. Open-weight models vary more, from 128K native on Granite 4.2 to 1,048,576 on Kimi K3.1 Pro, and the window you can serve in practice depends on your hardware.

Related: Open Weights in 2026: Chinese Labs Ship the Frontier, the Licenses Grew Clauses, and Brussels Started Enforcing, Running LLMs Locally in 2026: Ollama vs llama.cpp vs LM Studio vs MLX and LLM Inference in 2026: Prefill Is a Different Machine Now, and the KV Cache Has a Disk Tier.