Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
ExplainerWearables and AI dominate a busy week in digital health research02.10AnalysisUS Order Pushes Grid Supply Chain Scrutiny Into Software Layer02.10AnalysisThe humanoid hand fight moves to a factory floor that is already full02.10AnalysisSuperfluid qubit design, a DOE roadmap and a universal gate set02.10NewsSteam Deck 2 chip rumors meet a hard datacenter bottleneck02.10NewsSony starts rolling out QSSR AI upscaling on the base PS502.10AnalysisSmall models move to the edge as Google locks down its biggest one02.10ExplainerShopify Canvas Puts an AI Chat at the Centre of Store Building02.10AnalysisShopify Canvas Lets Merchants Build Stores by Chat, as AI Agents Breach a Security Nonprofit02.10AnalysisScience Journal Turns to Women's Health as Funding Cuts Threaten the Field02.10
›

New AI model releases are outrunning their own training data

Ten of the 20 current AI models listed by the freshness tracker stale.jock.pl carry a training cutoff their lab actually publishes. Several of them launched months behind the date on their own release announcement.

Grace OkonkwoGrace OkonkwoAI, models and technology
AI & modelsExplainerGrace OkonkwoPublished: 27 September 20263 min readSources 1
New AI model releases are outrunning their own training data

The site went up on Hacker News on 16 September. It tracks two dates per model: the day the lab shipped it and the day it stopped reading. It lists 20 models across 8 labs, and the browser counts upward from each date live.

The gap is the story. OpenAI released GPT-6 Astra on 3 September 2026, and its training cutoff is 30 April 2026. Its sibling GPT-6 Sol shipped on 22 September with a cutoff of 20 April, and GPT-6 Luna arrived the same day with a cutoff of 18 May. Anthropic's Claude Opus 5.5, released 22 September, stops at June 2026. Claude Sonnet 5 came out on 30 June and stopped reading in January.

For half the list, the answer is unknown. Mistral AI's Large 3, Small 4 and Medium 3.5, Alibaba's Qwen3.8-Max and Qwen3.8-Flash, Meta's Muse Glimmer and Muse Spark 1.3, DeepSeek's V4-Pro and V4.1-Flash, and Google DeepMind's Gemini 3.8 Flash all show no established cutoff. The page is careful about what that means. A blank is what the checked vendor sources failed to establish, not proof the lab never published one.

Why the date matters more than the version number

The site's own framing is blunt. A model can ship in September and still stop reading in April. That leaves it five months behind on launch day, before anyone types a prompt into it.

Search tools do not fix that, according to the page. When a model searches, it reads a few pages, uses them in one answer and forgets. Open a new chat and the cutoff is back. The trigger decision also runs on the same weights that hold the stale fact, so misses land where the model is most confident. The page says it measured this over 2,000 calls across 16 models, each given a web search tool. Frontier models decided correctly almost every time. Weaker ones stated settled facts that had changed and searched the web for 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.

That is the site's own line, and it is the cleanest illustration of the problem. A model that does not know a monarch has died will not reliably decide to check.

The workaround is aimed at agents, not people

The page publishes the same data as models.json, with release date, published cutoff and a source link per model. It suggests pasting a short instruction block into an AGENTS.md or CLAUDE.md file so an agent fetches it before naming any model as current.

On the lab side, the published cutoffs are uneven. Five of the eight labs, Anthropic, Google DeepMind, Meta, OpenAI and xAI, have a published cutoff for at least one model in the list. Ten of the 20 models carry one. The rest do not.

The page also offers a crude self-test: ask the model for its own training cutoff. One that answers, or admits uncertainty, is behaving. One that invents a confident date has told you something. The site's warning is that a model is a poor source on models, so check the answer against the table rather than trusting it.

Comments 0

Sources

1
  1. 01Show HN: How Stale Is Your AI? Release age and training cutoff for 20 modelsEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Grace Okonkwo

Grace Okonkwo

AI, models and technology

Grace Okonkwo covers AI, models and technology for FLASH24, working from primary sources such as model cards, API documentation and benchmark papers rather than vendor summaries. She checks training data provenance, evaluation conditions and reported scores against the underlying datasets before any figure reaches print. She interviews researchers and engineers directly, tracks release calendars from major labs, and compares successive model versions on the same tests. Her own self-hosting, home-network and documentation-reading habits feed straight into that desk, since she tests tools on her own hardware first. She does not publish benchmark claims without a reproducible method.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.

Previous articleChinese open models take majority share on OpenRouter and Vercel as Washington opens probesNext articleOpen weights, open source and one model that memorised Harry Potter