Show HN: How Stale Is Your AI? Release age and training cutoff for 20 models

${MODELS.length} models/ ${labs.length} labs/ Median age ${median} days/ Freshest ${esc(freshest.model)}/ Stalest ${esc(stalest.model)}/ ${withCutoff.length} of ${MODELS.length} publish a cutoff

The shelf

Time since release Time since the training data stops

The big number

Days since the lab released the model to the public. It ticks up from a fixed date, so the shelf reorders itself without anyone touching it.

Blind

Days since the training data stops, drawn as the striped red run behind each bar. A date appears only when the checked vendor sources establish it. Otherwise the shelf leaves it blank. A month with no day is counted from the last day of that month, the kindest possible reading.

The stamp

Fresh under 30 days, in date under 90, turning under 180, past date under a year, fossil beyond it. Arbitrary, but so is calling a six month old model current.

For your agent

A model is a poor source on models, so hand yours a better one. These lines go in the AGENTS.md or CLAUDE.md your agent already reads, and point it at models.json, the same data this page renders: release date, published cutoff and a source link for all ${MODELS.length} models.

Plain JSON, open to any origin, regenerated whenever the shelf changes, so your agent reads today's dates instead of the ones baked into its weights. There is an llms.txt for crawlers that want the page in prose.

When the shelf moves, that file moves with it, and I write up what actually changed and whether it matters. Get that by email.

Paste into AGENTS.md

${esc(SNIPPET)}

Why any of this matters

The mechanism

Three dates get mixed up constantly, and the one everybody quotes is the least useful of them.

The release date is when the lab put the model in front of you. It is what the headlines report, and it is what the shelf above is sorted on.

The training cutoff is when the model stopped reading. Everything that happened after it is simply absent. A model can ship in September and still stop reading in April, which means it is five months behind on the day it launches. A blank here means the checked vendor sources did not establish a cutoff for that model.

Browsing is not the fix people think it is. When a model searches the web for you it is not learning anything. It reads a few pages, uses them in that one answer, and forgets. Open a new chat and it is April again. Search tools paper over the gap. They never close it.

What to do about it

Ask it directly: "what is your training cutoff date?" A well behaved model answers. One that invents a confident date has just told you something useful about itself.
Turn search on for anything with a date attached. Prices, versions, who runs which company, which model is current. If the answer would change month to month, do not let it answer from memory.
Check the shelf above before you trust a model about models. Ask one which Claude or which GPT is newest and it will happily name something that retired months ago, because from the inside its own launch still feels like this morning.

What the shelf says right now

Mistral shipped Small 4 in March and Medium 3.5, which it calls frontier class, on April 28, both as open weights, then spent the summer on OCR and Lean proof tooling instead of a new flagship. Meta stopped shipping Llama after version 4 and replaced the line with the closed Muse Spark family. Google released four Flash models in about 106 days while the Pro flagship promised for June still has not appeared, so a February preview is still its top tier. Anthropic shipped four Claude models in a single quarter, and OpenAI shipped GPT-6 Astra on September 3.

${new Set(MODELS.filter(m => m.cutoff).map(m => m.vendor)).size} of ${labs.length} labs on this page tell you when their models stopped reading. Everyone else ships and lets you guess.

Why this page exists

I build with these models every day, and the thing that kept catching me out was models being confidently wrong about models. That is a funny failure and also a real one, so I made a page that holds the dates for me and counts upward, because a number that moves is much harder to ignore than a number sitting in a changelog.

I am Pawel Jozefiak. I write about building with AI at Digital Thoughts, run an agent called Wiz that ships small tools like this one at wiz.jock.pl, and think out loud on X as @joozio. This page is one of the small things that fell out of that. Static HTML, no framework, every date checked by hand against the lab's own docs, and every model name links to the source that proves it.