45 days of model releases, Oct 5 to Nov 19: the dated ledger and the four things that changed
I stopped trying to follow model launches one by one somewhere around the middle of October. BenchLM's tracker lists 47 releases for October and 23 for November so far, and most of them are speech, embedding, decision or translation models that never reach a headline. So I did the boring thing instead: one dated table for the 45 days from October 5 to November 19, restricted to the releases that can change a routing decision, and then a short list of what shifted underneath.
The ledger
Prices are list rates in dollars per million tokens, input then output, as published by the vendor or reported by the trackers I cross-checked. Where a model has no public price I say so.
| Date | Model | Lab | Access | Price in / out |
|---|---|---|---|---|
| Oct 6 | Mistral Large 4 | Mistral | API preview, weights on Oct 29 | 0.70 / 2.10 |
| Oct 6 | Decisions API on GPT-6 Luna | OpenAI | Public beta | 0.10 input |
| Oct 7 | Claude Haiku 5.5 | Anthropic | Generally available | 0.10 / 0.50 up to 100k |
| Oct 8 | Mellum2.1 | JetBrains | Open weights, Apache 2.0 | Self-hosted |
| Oct 12 | Muse Wonder | Meta | API and Muse Code | 4 / 20 |
| Oct 13 | Gemini 4 Argon | Generally available | 2 / 10 intro | |
| Oct 15 | Muse Spark Open | Meta | Open weights, Muse Model License | Self-hosted |
| Oct 16 | Beam 1.0 | Reflection | Open weights | Self-hosted |
| Oct 17 | DeepSeek V4.1 Flash 1017 | DeepSeek | API swap, MIT weights | Unchanged |
| Oct 23 | Kimi K3.1 Pro | Moonshot | API, weights followed in November | 2 / 10 |
| Oct 26 | GPT-6.1 Terra | OpenAI | Released | 1 / 5 |
| Oct 28 | Gemini 4 Carbon | Cyber defenders first, developers Nov 10 | 2 / 10 intro | |
| Oct 30 | GPT-6 Rosalind | OpenAI | Approved organisations | Not public |
| Nov 4 | Hy4 | Tencent | Open weights | 0.83 / 2.50 on OpenRouter |
| Nov 5 | GPT-6 Cyber | OpenAI | Daybreak Red only | Not public |
| Nov 9 | Claude Fable 5.2 | Anthropic | Generally available | 8 / 40 |
| Nov 9 | Claude Mythos 5.2 | Anthropic | Verified organisations | Gated |
| Nov 9 | Muse Ember | Meta | API | 0.10 / 0.50 up to 200K |
| Nov 11 | Gemini Decision 1 | Gemini API | 0.10 input | |
| Nov 12 | Gemini 4 Hydrogen | Generally available, free tier | 1.50 / 7.50 | |
| Nov 16 | Qwen4 Max | Alibaba | API | 1.50 / 4.50 |
| Nov 16 | Hy4 flash | Tencent | API | 0.07 / 0.281 |
| Nov 17 | GPT-6.2 Astra and Sol | OpenAI | Generally available | 8 / 40 and 2 / 10 |
| Nov 17 | MAI Code 2 and Code 2 Flash | Microsoft | GitHub Copilot | No list price published |
| Nov 18 | Gemini 4 Neon | Generally available, paid tier | 3 / 15 |
Three dates in that table are softer than they look. Gemini 4 Carbon was announced on October 28 for a subset of cyber defenders and only reached developers on November 10. Mistral Large 4 was a hosted preview on October 6 and became a downloadable model on October 29. Kimi K3.1 Pro is dated October 23 by some trackers and October 24 by others. If you automate anything off a release date, decide which of those events you mean.
Change one: the frontier refreshed in ten days and landed on one price
Anthropic shipped Fable 5.2 on November 9, OpenAI shipped GPT-6.2 Astra on November 17 and Google shipped Gemini 4 Neon on November 18. Fable 5.2 and Astra both list at $8 and $40. One step down, GPT-6.2 Sol sits at $2 and $10 and Neon at $3 and $15. I wrote in September that Astra and Fable had converged on price. They have now converged twice in a row, 20% lower than last time, which tells me list price has stopped being the variable that separates them. Tokens consumed per finished task is.
Change two: gated models are a product line now
Four entries in the ledger are models most readers cannot call, or could not call at launch: Mythos 5.2, GPT-6 Cyber, GPT-6 Rosalind, and for thirteen days Carbon. Google has also said a separate Neon for cyber is going to vetted defenders. A year ago a restricted release was an event. In this window all three large labs did it.
Change three: open weights reached a trillion parameters, with fine print
Mistral Large 4 is roughly 1T parameters with 52B active. Muse Spark Open is 816B with 42B active. Kimi K3.1 Pro is a 2.8T backbone. All three are downloadable, and none of them is under a plain permissive license the way the 12B Mellum2.1 is. Mistral's license is MIT plus one clause about hosting, Meta's has what it calls limited commercial terms, and for Kimi I still have not seen the full license text reproduced by a source I trust.
Change four: a ten-cent tier and models that never write a word
Haiku 5.5 and Muse Ember both list at $0.10 input. Hy4 flash undercuts them at $0.07. Next to them sits a category that did not exist in the summer: decision models that score a closed list of options and return probabilities instead of text. OpenAI's Decisions API went to public beta on October 6, Microsoft followed within days at $0.042 per million input tokens, and Google joined on November 11.
A release date tells you when a vendor pressed publish. It does not tell you when you could call the model, download it, or afford it.
What I am changing
I am re-running my own task set against Sol, Neon and Hydrogen before I touch any routing rule, because every cross-vendor benchmark number in this window comes from a launch post written by one of the competitors. I am moving classification and routing calls off general models and onto the ten-cent tier or a decision model, since that is where the measurable saving is. And I am putting December 1 in the calendar, because that is the day Google's introductory prices for Argon and Carbon are scheduled to double.
The limitation of a ledger like this is that it is already stale. It also leaves out roughly fifty speech, embedding, image and translation releases on purpose. Treat it as a map of what to test and check the vendor page before you quote a number from it.