Knowledge cutoff

A knowledge cutoff is the date after which a model was not trained on anything. Material published later exists nowhere in its weights. The model can still discuss it, but only if something puts the material in front of it at answer time.

Vendors publish the date because it is a capability boundary, not a footnote. Anthropic's model comparison table gives one row per model and, as fetched on 4 October 2026, listed reliable knowledge cutoffs of June 2026 for Claude Fable 5.1, Claude Opus 5.5 and Claude Sonnet 5.5, and February 2025 for Claude Haiku 4.5. A month earlier the same table listed Claude Opus 5 at May 2026 and Claude Sonnet 5 at January 2026; the lineup moves with each release.

Why does "reliable" matter in "reliable knowledge cutoff"?

The wording is not hedging. A training corpus does not end cleanly on a date: coverage of the last months before collection is thin and uneven, because the web had not finished writing about that period. A page published in the final weeks may appear once, without the commentary, corrections and repetition that make a fact stick.

So a reliable cutoff is the date up to which the vendor will say the knowledge holds, which is earlier than the newest document in the corpus. Read the published date as the edge of dependable knowledge, not the edge of any knowledge.

Do models from the same vendor share a cutoff?

No. Anthropic's table makes a point worth keeping: models available on the same day, from the same vendor, can have cutoffs more than a year apart. A smaller or older model is not only less capable, it is less current.

That matters when an assistant routes work across models to control cost. The cheap model handling a routine question may be working from a materially older world than the expensive one, and nothing in the answer says so.

How do answers get past the cutoff?

By fetching. A model with a search or retrieval tool reads current pages and writes from them, which is what grounding means: Google describes its search tool as letting a model cite verifiable sources beyond its knowledge cutoff. Retrieval-augmented generation does the same over a private corpus.

Our own server logs show which crawlers do the fetching. attensira.com counts every request from a self-identified AI crawler at the edge and classes it by purpose, and in the 30 days to 4 October 2026 the operators split cleanly in two, as the figures at the top of this page show. The answer engines' agents came almost entirely to index or to answer now: the paths that carry a page published last week past any cutoff. Meta's and ByteDance's crawlers came only to collect training data, which can reach a model at its next training run at the earliest, behind a new cutoff. The request volume per operator is on the pay per crawl page. These are user-agent claims, which can be spoofed; see training versus search crawlers for how the agents differ.

What does a cutoff not protect you from?

Confident staleness, from two directions.

From the weights. A parametric answer carries no timestamp. A discontinued product, a renamed company or an old price is stated in the present tense. That is a different failure from a hallucination: the fact was true once and the model is not inventing it, but it reads the same to the buyer.

From your own old pages. Retrieval reads whatever it can reach. The ChatGPT answer to "what is Attensira" stored in our own workspace on 4 October 2026, in the band below, describes Attensira as aimed mainly at SaaS and developer-tool companies, a narrower audience than the site describes today, and cites attensira.com. Nothing was wrong with the model. It fetched attensira.com and other pages that describe us, and wrote from what they said that day. Which cited page carries the narrower audience is the first thing to check, and it is a page, not the weights.

What should a brand do about knowledge cutoffs?

• Treat recent changes as at risk until retrieval finds them. A rebrand, a pricing change or a retired plan lives on the web before it lives in a model. The gap closes through pages an agent can fetch, not by waiting for the next training run.
• Retire or date old pages. A stale comparison or feature list is a live document to a retrieval system. Delete it, or date its claims in the text ("as of March 2026"), which survives being lifted into a passage.
• Keep crawlers able to fetch you. Blocking search and user-triggered agents leaves the weights as the only route, and the weights are the stale route.
• Never assume the model you tested is the one answering. Assistants route between models and versions. Sample repeatedly rather than trusting one transcript.

How does Attensira catch outdated claims?

By asking the questions again every day and noticing when the answer is out of date, then showing you which source kept it alive. The stale version usually has an address. Our own example above is one: the answer that narrows Attensira's audience cites attensira.com, so the first place to look is the pages it cited, not the next model update. Attensira checks the factual statements answers make about you against the facts you confirmed in your brand profile and flags a contradiction with the quote and the sources the answer cited. A cited page of yours means the fix is on your site; no cited source points to the model's memory, which only fresh, fetchable pages displace. Facts you have not confirmed are never used, so a claim with nothing to check it against is reported as exactly that.