New AI models ship with months-old training data, tracker shows

The page is called How stale is your AI. It does one thing: it puts two dates next to 20 model names and lets them count upward. The release date is when the lab shipped the model. The training cutoff is when it stopped reading. According to the site, 10 of the 20 models listed carry a cutoff their lab actually publishes.

That is the whole benchmark argument in miniature. Model launches get graded on benchmarks run at release, but everything after the cutoff is simply absent from what the model knows. The tracker's own example: GPT-6 Astra was released on 3 September 2026 and its training data stops at 30 April 2026. On launch day it was already four months behind.

What the dates show

Take the entries where both dates exist. Llama 4 shipped 5 April 2025 with an August 2024 cutoff. Claude Haiku 4.5 shipped 15 October 2025 with July 2025. Gemini 3.1 Pro shipped 19 February 2026 with January 2025, a gap of more than a year. Claude Sonnet 5 (30 June 2026, January 2026), Claude Fable 5.1 (1 September 2026, June 2026), GPT-6 Astra (3 September 2026, 30 April 2026), Grok 4.7 (21 September 2026, May 2026), Claude Opus 5.5 (22 September 2026, June 2026), GPT-6 Sol (22 September 2026, 20 April 2026) and GPT-6 Luna (22 September 2026, 18 May 2026) fill out the published set.

The other ten rows are blank. Mistral Large 3, Mistral Small 4, Mistral Medium 3.5, Qwen3.8-Max, Muse Glimmer, DeepSeek V4-Pro, Qwen3.8-Flash, Gemini 3.8 Flash, Muse Spark 1.3 and DeepSeek V4.1-Flash all show "Not established." The site is careful about what that means: a blank means the checked vendor sources did not establish a cutoff for that model, not proof the lab never published one. Five of the eight labs, Anthropic, Google DeepMind, Meta, OpenAI and xAI, have a published cutoff for at least one model on the page.

The training cutoff is the date a model stopped reading. Everything that happened after it is simply absent from what the model knows. A model can ship in September and still stop reading in April, which means it is five months behind on the day it launches.

The page's most interesting claim is not about dates at all. It is about what happens when you hand a model a search button and let it decide whether to use it. The author reports 2000 calls across 16 models, each holding a web search tool. Frontier models, in that run, decided correctly almost every time. Weaker ones answered settled questions from memory after the answer had changed, and searched the web for things like the boiling point of water. In one framing on the page, 16 models were asked who the king of Norway is, and five named a dead man.

Read that carefully. The search decision runs on the same weights that hold the stale fact, so the misses land exactly where the model feels most certain. Search tools paper over the gap. They do not close it. Open a new chat and the model is back at its cutoff.

The awkward part for benchmarks

This matters for how release benchmarks get read. A model can top a leaderboard in September on questions whose answers were settled before April, and no one running the eval is checking the cutoff date on the model card. If a lab does not publish the cutoff, the tracker's blank is the honest answer: nobody outside the lab can measure the gap.

There is a second-order problem too. A model is a poor source on models. Ask it directly what its training cutoff is and a well behaved one answers or says it is not sure, per the page. One that invents a confident date has told you something useful about itself. Release dates, version numbers and which model is current all stop at the cutoff, and the wrong answer still sounds confident.

The suggested fix is small and practical. Put a few lines in the AGENTS.md or CLAUDE.md file an agent already reads, telling it to fetch the site's models.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 whenever the page is. The page hosts the wording, ready to paste.

None of this is a benchmark in the usual sense. There is no score, no leaderboard and no lab-submitted run. It is a date table with a JSON export and an argument attached: that the second date on a model card deserves the same scrutiny as the first.

For anyone comparing two models released the same week, the tracker adds a question that most launch coverage skips. When did each one stop reading, and did the lab bother to say?