Two Dates Decide How Stale an AI Model Is, and Only 10 of 20 Labs Publish Both

The page, stale.jock.pl, went up as a Show HN submission on 16 September. Its premise is narrow. A model can ship in September and still have stopped reading in April, so the gap is baked in on launch day, before anyone types a prompt. The same data is offered as models.json for agents to fetch.

Five of the eight labs publish a cutoff for at least one model listed: Anthropic, Google DeepMind, Meta, OpenAI and xAI. A blank in the table means the vendor sources checked did not establish a cutoff, the page says, not proof that the lab has never published one. Mistral, Alibaba and DeepSeek appear in the table with no established cutoff for the models shown.

The gap, model by model

OpenAI's GPT-6 Astra was released on 3 September 2026 with training data stopping at 30 April 2026, a four-month gap. GPT-6 Sol, released 22 September, stops at 20 April. GPT-6 Luna, same release date, stops at 18 May. Anthropic's Claude Opus 5.5 shipped 22 September with a June 2026 cutoff; Claude Fable 5.1 shipped 1 September, also June 2026. Meta's Llama 4, released 5 April 2025, has a cutoff of August 2024.

The longest lag in the table belongs to Gemini 3.1 Pro: released 19 February 2026, cutoff January 2025, thirteen months. Mistral Large 3, Mistral Small 4, Mistral Medium 3.5, Qwen3.8-Max, Qwen3.8-Flash, Muse Glimmer, Muse Spark 1.3, DeepSeek V4-Pro and DeepSeek V4.1-Flash are all listed without a published cutoff, so no gap can be computed for them from this source.

The page argues that search tools do not fix the problem. When a model searches, it reads a few pages, uses them in one answer and forgets. Open a new chat and it is April again. The search also has to be triggered by the model, using the same weights that hold the stale fact. The site claims a measurement of over 2,000 calls across 16 models, each given a web search tool. Frontier models decided correctly almost every time. Weaker ones answered settled questions from memory after the answer had changed, and searched the web for the boiling point of water. The page also says that when 16 models were asked who the king of Norway is, five named a dead man.

Why the second number is missing

There is a commercial logic to the silence. A release date is a marketing event and gets a launch post. A cutoff is an admission about what the model does not know, and a lab that publishes it hands competitors and customers a way to measure the gap on day one. Ten out of twenty is the result.

The page suggests a workaround for people building agents: put a few lines in the AGENTS.md or CLAUDE.md file the agent already reads, telling it to fetch the JSON before naming any model, version or date as current, and to treat anything with a date attached as unverified until checked. The export carries release date, published cutoff and a source link per model, and is regenerated with the page, so the agent reads today's dates rather than the ones in its weights. The wording is on the page, ready to paste.

The obvious limit is that the page is a model's account of models, and it says so itself. Its advice for checking a single model is to ask it directly what its training cutoff is. A well behaved model answers or says it is not sure. One that invents a confident date has told you something useful about itself. Check the answer against the table before trusting it.

No independent audit of the 20 dates is offered on the page beyond the linked lab documents, one per model. Anyone relying on the numbers for a shipping decision should follow those links, because the blank cells are the ones most likely to move first.