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 rather than a footnote. Anthropic's model comparison table gives one row per model, and — as fetched on 4 September 2026 — lists reliable knowledge cutoffs of June 2026 for Claude Fable 5.1, May 2026 for Claude Opus 5, January 2026 for Claude Sonnet 5, and February 2025 for Claude Haiku 4.5.

Why "reliable" is doing work in that phrase

The wording is not marketing hedging. A training corpus does not end cleanly on a date: coverage of the last few months before collection is thin and uneven, because the web had not finished writing about that period yet. 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 is prepared to say the knowledge holds, which is earlier than the newest document in the corpus. Treating the published date as the boundary of dependable knowledge, rather than the boundary of any knowledge, is the correct reading.

Different models, different dates

The table above makes a point worth internalising: models available at the same moment, from the same vendor, can have cutoffs more than a year apart. A smaller or older model is not merely less capable, it is less current.

That matters when an application routes work across a model roster 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 output will say so.

What a cutoff does not prevent

It does not prevent an answer about recent events. A model with a search or retrieval tool fetches current pages and writes from them, which is precisely what grounding means — Google describes its search tool as letting the model cite verifiable sources beyond its knowledge cutoff. The same is true of retrieval-augmented generation over a private corpus.

It also does not prevent confident wrongness. This is the part that hurts. A parametric answer carries no timestamp: the model does not experience its knowledge as old, so a discontinued product, a renamed company or a superseded price is stated in the present tense with no signal that it expired. That is a different failure from a hallucination — the fact was true and the model is not inventing it — but it reads identically to the person receiving it.

What follows for a brand

Three practical consequences.

Anything you changed recently is a risk surface until retrieval finds it. A rebrand, a pricing change, a discontinued plan or a new positioning statement lives in the world before it lives in a model. The gap is closed by pages an agent can fetch, not by waiting for the next training run.

Old pages you left up are still being read. A stale comparison page or an outdated feature list is a live document to a retrieval system, and it will be treated as current unless it says otherwise. Dating claims in the text — "as of March 2026" — is the cheapest correction available, and it survives being lifted into a chunk.

Never assume the model you tested is the model answering. Assistants route between models and versions, and a check performed against one may not describe what another says. Where the answer matters, sample repeatedly rather than concluding from a single transcript.

Finally, expect these dates to move. The roster in this page's source changes with each release, which is why the figures above are stated as of the day the page was fetched rather than as standing fact.