AI model release tracker: date, context window and licence

Twelve releases re-read on 21 November 2026. A blank means the lab’s own page did not say it.

This is a lookup table, not a ranking and not a buying guide. It records twelve major model releases whose own announcements were re-read on 21 November 2026. Each row keeps four fields that a citation actually needs: who released the model, the date on the lab’s page, the context window if that page states one, and the licence if that page names one. Prices and benchmark scores are left out on purpose. Both move, and a number copied from a launch post goes stale faster than the licence does.

A cell that says “Not stated” is not a zero and it is not a guess. It means the page linked from that row did not give the field when it was read. A second page is linked only when the date or the context window comes from somewhere other than the announcement, and that second page is named in the note. The wording in the licence column is the lab’s description. It is not this desk’s reading of whether the terms meet the Open Source Initiative’s definition of open-source AI, and it is not legal advice.

The rows sit next to the reporting that already quotes these posts. OpenAI, Anthropic, Google and DeepMind, and the open-weight hub collect the stories. What is a large language model? is the training term behind most of the product names. A world model, a speech model, or a weights file this desk has covered but did not re-open on 21 November is not in the table. Absence from the table is not a judgment that the release was small.

The same twelve rows are in a CSV download. On a phone the table scrolls sideways. The short citation is under the table.

Major AI model releases re-read on 21 November 2026, newest first. Context window and licence are quoted from the linked lab page.
Model Maker Date Context window Licence Announcement
Gemini 4 Neon Google The announcement is dated 18 November 2026 and says the model begins rolling out that day. 1 million tokens The announcement says a 1 million-token context window. Not stated The announcement describes API and app access. It does not name a licence or say the weights are released. Introducing Gemini 4 Neon Google’s Gemini 4 Neon launch
GPT-6.2 Astra and GPT-6.2 Sol OpenAI The product post says “starting today” and does not print a calendar date. The system card is dated Published 17 November 2026. Not stated Neither the product post nor the system card states a context window. Not stated The product post says developers can call gpt-6.2-astra and gpt-6.2-sol in the API. It does not name a licence or say the weights are released. Introducing GPT-6.2 Astra and Sol GPT-6.2 system card OpenAI ships GPT-6.2 Astra and Sol
Gemini 4 Hydrogen Google The announcement is dated 12 November 2026. The same page says it was last updated 16 November 2026. 1 million input tokens; up to 1 million output tokens The announcement says the price is flat across a 1 million-token context window, and that the model pairs 1 million tokens of input with up to 1 million output tokens. Not stated The announcement describes API, app, and Search access. It does not name a licence or say the weights are released. Introducing Gemini 4 Hydrogen Google launches Gemini 4 Hydrogen
Qwen4-Max-Preview Alibaba Cloud The model page, last updated 16 November 2026, says the preview runs from 10:00 on 11 November to 24:00 on 25 November 2026 (UTC+8). The same page also says a production Qwen4-Max is available and recommended. This row is the preview the page dates. 1 million tokens The model page lists a context length of 1 million tokens and a maximum output of 393,216 tokens, shared by thinking and the answer. Not stated The page describes hosted Token Plan access. It does not name a licence or say the weights are released. qwen4-max-preview Alibaba opens a Qwen4-Max preview
Praxis-1 Runway The announcement is dated 11 November 2026. Not stated Not stated. The announcement describes a 27-billion-parameter action model and does not give a text context window. Apache 2.0 Runway says the weights, inference code, and fine-tuning code are released under Apache 2.0, including for commercial use. Praxis-1 Is Now Available Runway releases Praxis-1
Claude Fable 5.2 and Claude Mythos 5.2 Anthropic The announcement is dated 9 November 2026. An edit note on the page is dated 11 November. Not stated The announcement does not state a context window. Not stated The announcement says Fable and Mythos are the same underlying model. It says Fable 5.2 is on the Claude Platform as claude-fable-5-2, and that Mythos 5.2 access is through verification programs. It does not name a licence or say the weights are released. Introducing Claude Fable 5.2 and Claude Mythos 5.2 Anthropic launches Fable 5.2 and Mythos 5.2
GPT-6 Cyber OpenAI The announcement is dated 5 November 2026. Not stated The announcement does not state a context window. Not stated The announcement says the model is initially available through the API and Codex to approved Daybreak Red customers. It does not name a licence or say the weights are released. GPT-6 Cyber OpenAI introduces GPT-6 Cyber
Kimi K3.1 Pro Moonshot AI Moonshot’s post on X is dated 4 November 2026 and says the weights are being released. The model card itself does not print a calendar date. 1,048,576 tokens The model card lists a context length of 1,048,576 and also calls it a 1-million-token context window. Kimi K3.1 Pro License The model card says the model is released under the Kimi K3.1 Pro License. The card section read on 21 November 2026 does not paste the licence terms. Kimi K3.1 Pro model card Moonshot’s release post Moonshot releases Kimi K3.1 Pro weights
Tencent Hy4 Tencent The announcement is dated 4 November 2026. This row records that page only. Not stated The announcement discusses long-context comprehension and does not give a token count. Not stated The headline calls Hy4 open-source. The same post says Tencent Hy plans to release the weights and system cards in the coming days. It does not name a licence. Tencent launches Hy4 Tencent releases Hy4
Mistral Large 4 (weights) Mistral The weights announcement carries a published time of 29 October 2026. The post says “today” and does not repeat the calendar date in the body. 1 million tokens Mistral’s model page lists the context as 1M. Read 21 November 2026. Mistral Ecosystem Open-Weight licence Mistral calls it a modified MIT licence: free to use, modify, fine-tune, and commercialize, except that a business offering the model as a service needs a separate commercial licence from Mistral. Mistral Large 4 is now open weight Mistral Large 4 model page Mistral releases the Mistral Large 4 weights
Gemini 4 Carbon Google The announcement is dated 28 October 2026. A later availability post was also read and does not add a context window or a licence. Not stated Not stated on the 28 October announcement or on the later availability post. Not stated The announcement describes a limited rollout, including a cyber-defender program, and does not name a licence or say the weights are released. Introducing Gemini 4 Carbon Gemini 4 Carbon is now available Google’s Gemini 4 Carbon announcement
GPT-6.1 Terra OpenAI The system card says Published 26 October 2026. The product post does not print a calendar date. Not stated Neither the product post nor the system card states a context window. Not stated The product post says the API name is gpt-6.1-terra. It does not name a licence or say the weights are released. Introducing GPT-6.1 Terra GPT-6.1 Terra system card OpenAI releases GPT-6.1 Terra

What the blanks show

The useful comparison is not which model scored higher. It is which labs published the two fields a reader needs in order to cite the release, and which left them off the page.

Related reporting

Google’s Hydrogen post of 12 November and Neon post of 18 November each state a 1 million token window. Hydrogen also states an output cap of up to 1 million tokens. Carbon’s announcement of 28 October does not, and the later post that says Carbon is available does not add one. None of the three Google posts names a licence or says the weights can be downloaded.

OpenAI’s Terra, Cyber, and GPT-6.2 posts likewise do not state a context window or a licence name. Two of the dates are not on the product post at all. Terra’s system card says it was published on 26 October 2026. The GPT-6.2 system card says it was published on 17 November 2026. The Cyber announcement itself is dated 5 November 2026, and it limits first access to approved Daybreak Red customers. Those are access facts from OpenAI’s pages, not a finding about what the model can do.

Anthropic’s 9 November announcement says Claude Fable 5.2 and Claude Mythos 5.2 are the same underlying model, with Fable on general sale and Mythos behind verification programs. It names the API id claude-fable-5-2. It does not state a context window or a licence, and it does not say the weights are released.

The rows that do name a licence are the ones that also hand out a file. Mistral’s 29 October weights post calls the terms a modified MIT licence, the Mistral Ecosystem Open-Weight licence, and it adds one condition: a business that offers the model as a service needs a separate commercial licence. Moonshot’s model card says Kimi K3.1 Pro is under the Kimi K3.1 Pro License and lists a context length of 1,048,576 tokens. Runway’s 11 November post says Praxis-1’s weights and the code to run and fine-tune them are under Apache 2.0, including commercial use. The page does not give a text context window. It describes a 27-billion-parameter action model, which is a different specification.

Two rows are easy to misread from the headline alone. Tencent’s 4 November Hy4 post uses “open-source” in the headline and then says the weights and system cards are planned for the coming days. It does not name a licence and it does not give a token count. Alibaba’s Qwen4-Max-Preview page, last updated 16 November, lists a context length of 1 million tokens and does not name a licence. It dates the preview from 11 November through 25 November 2026 (UTC+8), and the same page says a production Qwen4-Max is available and recommended. This table does not turn that sentence into a second row, because the page does not give the production model its own date, context window, or licence.

Cite this

AI Headline Hub, “AI model release tracker: date, context window and licence,” 21 November 2026, https://aiheadlinehub.com/model-releases/. Context windows and licences are the labs’ own statements on the linked pages, checked that day. The desk did not re-run the models. The CSV has the same rows.

What was left out

A launch price is not in the table even when the lab printed one. The editorial policy on this site is to label a vendor figure as the vendor’s, and prices on these pages have already been revised inside a single product family. A score on someone else’s index is left out for the same reason. Where a story on this site reports one, the link in the last column goes to that story and to the lab post it cites.

The table is also not a census. It stops at releases whose primary page was open on 21 November 2026 and whose maker, date, and announcement could be tied to that page. Genie 4, NVIDIA’s Nemotron Transcribe, Pika H3 Flex, MiniMax H3-Image, and Microsoft’s MAI weights plan are covered as stories and are not rows here. The next edit should add a row only when a new lab page has been read, and it should change a cell only when that page has changed. The date at the top of this page moves when that happens, not when the site is rebuilt.

Maintained by AI Headline Hub. General information, not purchasing, security, or legal advice. See the disclaimer.

Newest related stories