What a Model's Two Dates Say About How Current It Really Is

The page is called "How stale is your AI?" and it went up on Hacker News on 16 September. It holds two dates for 20 models across eight labs, release date and training cutoff, and counts upward from each one live. The counters need JavaScript. The dates do not.

The distinction matters more than it sounds. The release date is when the lab shipped the model, and it is what the launch coverage reports. The training cutoff is the date the model stopped reading. A model can ship in September and still stop reading in April. That puts it five months behind on the day it launches.

Take GPT-6 Astra, which OpenAI released on 3 September 2026 with a cutoff of 30 April 2026. Or Claude Opus 5.5, released 22 September 2026, cutoff June 2026. The tracker's data lists Llama 4 from Meta at 5 April 2025 with a cutoff of August 2024, and Gemini 3.1 Pro from Google DeepMind at 19 February 2026 with a cutoff of January 2025.

Ten of the 20 models carry a cutoff the lab actually publishes. Five of the eight labs have published a cutoff for at least one model on the page: Anthropic, Google DeepMind, Meta, OpenAI and xAI. A blank entry means the checked vendor sources did not establish a cutoff for that model. It is not proof the lab never published one, the page says.

Why search does not fix it

Search tools paper over the gap. They never close it, the page argues, because the model itself triggers the search, using the same weights that hold the stale fact. When a model searches, it reads a few pages, uses them in that one answer and forgets. Open a new chat and it is April again.

The site backs that with a test: more than 2,000 calls across 16 models, each handed a web search tool. The frontier models decided correctly almost every time, the page says. Weaker ones stated settled facts that had changed without checking, and searched the web for things like the boiling point of water.

I gave 16 AI models a search button and asked who the king of Norway is. Five named a dead man.

The page also points to a second problem that is harder to test: a model is a poor source on models. Ask one for its own training cutoff. A well behaved model answers, or says it is not sure. One that invents a confident date has just told you something useful about itself.

Agents reading their own stale facts

The page ships a machine readable export, models.json, with release date, published cutoff and a source link for each of the 20 models. It also ships paste-ready wording for the AGENTS.md or CLAUDE.md file an agent already reads. That wording tells the agent to fetch the export before naming any model, version or date as current. The argument is simple: the agent's knowledge of models stops at its own cutoff, and a wrong answer still sounds confident.

The data is not the last word on which model is best. It is a check on one claim, currency, that benchmarks and launch posts rarely address. The page says the export is regenerated whenever the page is, so an agent reads today's dates rather than the ones baked into its weights.

Two caveats worth keeping. Ten models have no published cutoff on the page, so their staleness cannot be measured from this source alone. And the search test is the author's own, run at a size the page does not break down per model. Treat it as a signal, not a leaderboard.

What the page does establish is a habit: before quoting a model's knowledge as current, find the date it stopped reading. The launch date will not tell you.