Release date is not the same as training cutoff: how to read a new AI model

The tracker lives at stale.jock.pl and was posted to Hacker News on 16 September 2026. It lists 20 models from 8 labs. Two dates sit beside each one: the release date, and the training cutoff, the point at which the model stopped reading. Both counters tick upward live from those dates.

Eleven of the 20 rows show "Not established." That is not the same as "never published." The page says a blank means the checked vendor sources did not establish a cutoff for that model, so be careful reading the table as a claim about what labs do or do not disclose.

What the two dates actually measure

The release date is when the lab shipped the model. It is what headlines carry and what benchmark posts lead with. The training cutoff is when the model stopped reading. The distance between them is how far behind the model already was on launch day, before a single user typed anything into it. The page gives one clean example. GPT-6 Astra was released on 3 September 2026, and its training data stops at 30 April 2026. That is roughly four months of absence baked in at launch. The dates in the table come from lab documents, and each model name links to the document the date was taken from.

Some rows make the gap wider. Gemini 3.1 Pro, from Google DeepMind, is listed with a release of 19 February 2026 and a cutoff of January 2025, more than a year. Llama 4 from Meta shows a release of 5 April 2025 against an August 2024 cutoff. Anthropic's Claude Sonnet 5 shows 30 June 2026 against January 2026. Not every gap is large: Grok 4.7 from xAI, released 21 September 2026, carries a cutoff of May 2026, and Claude Opus 5.5, released 22 September 2026, carries June 2026.

Five labs, ten models, and a lot of blanks

According to the page, 5 of the 8 labs have a published cutoff for at least one model listed: Anthropic, Google DeepMind, Meta, OpenAI and xAI. Three labs, Alibaba, DeepSeek and Mistral AI, have no model on the page with an established cutoff. Alibaba's Qwen3.8-Max and Qwen3.8-Flash, DeepSeek's V4-Pro and V4.1-Flash, and Mistral's Large 3, Small 4 and Medium 3.5 are all listed as not established. That is a disclosure gap, not a performance claim. The tracker does not say those models are worse or older. It says the sources checked did not yield a cutoff date.

This matters for benchmark reading. A benchmark run is a snapshot of what a model can do on a task set, and task sets are written by people who know what happened up to some date. If a model's cutoff is older than the benchmark, some of what looks like reasoning may be recall. If the cutoff is newer, the benchmark may be measuring the model's ability to handle material that was in its training data.

The tracker does not make that argument directly. It makes a narrower one: you cannot evaluate currency without knowing the cutoff, and for half the models on the page the cutoff is not published.

The search button does not fix it

Web search is the usual answer to a stale model. The page pushes back on that, and it has numbers. Over 2000 calls across 16 models, each given a web search tool, frontier models decided correctly almost every time about whether to search, while weaker models stated settled facts that had changed without checking, and searched the web for things like the boiling point of water.

The search also has to be triggered by the model, using the same weights that hold the stale fact, so the misses land exactly where the model feels most certain.

That is the mechanism the tracker is pointing at. A search tool is only useful if the model recognises that it needs one, and that recognition runs on the same weights that hold the outdated belief. The page also notes that search does not update anything: the model reads a few pages, uses them in that answer, and forgets. Open a new chat and it is April again.

The page's own illustration is a king of Norway question put to 16 models given a search button. It says five named a dead man. That is one anecdote from a larger run, not a published benchmark with a methodology section, and it should be read that way.

What the tracker is actually for

The site ships a models.json export carrying the release date, published cutoff and a source link for all 20 models. It also ships a block of text intended to be pasted into an agent's instruction file, the AGENTS.md or CLAUDE.md that coding agents already read, telling the agent to fetch that JSON before naming any model, version or date as current. The argument for that is blunt and the page makes it: a model is a poor source on models. Ask a model for its own training cutoff and a well behaved one answers or says it is unsure. One that invents a confident date has told you something about itself.

For anyone reading a launch post this week, the practical version is short. Check whether the lab published a cutoff. Check how old it is against the release date. Treat a missing cutoff as missing information rather than a clean bill of health, and do not assume a search tool closes a gap that the model has to decide to open.

The tracker itself is a single page with a fixed list, updated when the shelf moves. It is not a benchmark and does not score models. It is a date table with counters, and the dates are the point.