Knowledge cutoff — the date a model stops learning
In one paragraph
A knowledge cutoff is the date after which nothing entered the model. It is published by the vendor, it differs per model rather than per brand, and it is softer than a date suggests: Anthropic gives two cutoffs per model, and Google describes one model as reliable to March 2026 in some domains and only to January 2025 in others.
what the model learned — and when it stopped
- years of books, sites and code
- last year’s prices and product names
- the news — up to a certain month
training ends — the model is finished
- yesterday’s headlines
- the new version of your tool
- today
Fourteen models against a date you pick. Their published cutoffs run from October 2023 to May 2026, and vendors do not all mean the same thing by the word: Anthropic gives Claude Haiku 4.5 a reliable knowledge cutoff of February 2025 and a training data cutoff of July 2025, five months apart.
What the date means
Training consumes a body of text and ends. What comes out is a fixed set of numbers, and the date that body of text stops is the cutoff. Nothing after it is inside the model, and running the model does not add anything: the next version does that.
Where it gets interesting is that vendors publish this date in three noticeably different ways.
Two dates, defined separately. Anthropic’s comparison table carries a “Reliable knowledge cutoff” and a “Training data cutoff”, with a footnote that defines them:
Reliable knowledge cutoff indicates the date through which a model's knowledge is most extensive and reliable. Training data cutoff is the broader date range of training data used.
For Claude Haiku 4.5 they are February 2025 and July 2025 — a five-month band the model has seen something about and is not underwritten for.
One date, hedged by topic. Google’s model card for Gemini 3.6 Flash:
The knowledge cutoff date for Gemini 3.6 Flash is March 2026 – users can expect updated information for some domains while in others they may experience the model's knowledge is limited to January 2025.
Fourteen months of spread, depending on the subject.
A single flat statement. OpenAI’s model pages give one line per model, such as the “Feb 16, 2026 knowledge cutoff” on the GPT-5.6 Sol page. Simple to read, and it discloses less than the other two.
Why it matters more than it looks
The cutoff is invisible from inside the conversation. A model has no clock and usually no date in the request, so it cannot notice that a question falls outside its range. The answer therefore arrives in exactly the same confident register as every other answer.
Nor is the boundary sharp in practice. Coverage of the last months before a cutoff is thinner than of earlier periods, which is what Anthropic’s two dates are trying to express. Treat the published date as the optimistic end of a slope rather than a wall.
Technical detail: how our dataset records the awkward cases
Two of the Gemini models in our dataset state no cutoff of their own. Their model cards defer known limitations to an earlier card, which defers to another, which gives January 2025. That date is inherited through a chain of references rather than stated for the model, and the dataset records it that way instead of quietly presenting it as first-hand.
Where a vendor publishes both a reliable and a training-data cutoff, we store the reliable one, because it is the date the vendor stands behind. The other is noted alongside it.
Common questions
- 1 Does the model learn from my conversations?
- Not into its weights. Those are fixed when training ends. What you type affects the request in front of it, and a memory feature may store notes and reinsert them later, but neither of those changes what the model knows.
- 2 Can I ask the model for its own cutoff?
- You can, and the answer is not a reliable source. It comes either from training text or from a system prompt written by the application, and both go stale. The vendor's model page is the one that is maintained.
- 3 Why do two models from the same vendor differ?
- Because a cutoff belongs to a training run, not to a company. In our dataset the OpenAI entries alone span October 2023 to February 2026, and which one you are talking to depends on the product and the plan.