Knowledge cutoff: what the AI doesn't know yet
A knowledge cutoff is the point at which an AI model stopped receiving new training data. It knows nothing about what happened afterwards, unless it looks the information up on the web.
Data as of October 22, 2026
Key figures
| Level | Beginner |
|---|
A snapshot of the world, taken on a specific date
A large language model learns from a huge amount of text: its training data. That collection stops on a given day. The model then « freezes » its view of the world at that point, a bit like a printed encyclopedia that doesn’t update itself.
A concrete example: you ask an assistant for the current price of a subscription, the name of a newly appointed official or last weekend’s game score. If it doesn’t check the web, it answers with what it knew at its cutoff, sometimes with great confidence. The answer looks right, but it’s out of date.
Two dates that don’t mean quite the same thing
Some AI makers distinguish between two notions. Anthropic, for instance, publishes a « reliable knowledge cutoff » for its Claude models, up to which the model’s knowledge is most complete, and a broader « training data cutoff ». In between, the model has seen some information, but only partially.
Another point: the cutoff isn’t the release date. A model may be published several months after its data collection ended. The site’s model list helps you place each model in time.
Why it matters to you
- News, prices, laws, software versions: these are the most exposed topics. Always check how recent the information is.
- The model may fill gaps by making things up. Asked about an event after its cutoff, it can produce a plausible but false answer, known as a hallucination.
- There’s a workaround: use an assistant that can search the web, or give it an up-to-date document yourself. That’s the idea behind RAG: feeding the model recent sources at answer time.
Common misconceptions
- « AI is always connected to the internet. » Not always: without search turned on, it answers from memory.
- « If the model is new, its knowledge is too. » No, there’s always a lag between cutoff and release.
- « The model knows what it doesn’t know. » Rarely: it’s poor at flagging the limits of its knowledge, which is why checking matters.
Related terms
- Training data · The examples (text, images, audio and more) an AI model learns from; they shape what the model knows and where it goes wrong.
- Pre-training · The first training phase of a language model: it learns to predict the next piece of text across a huge corpus, with no specific task.
- AI search · A search where an AI reads web pages for you and writes a summarized answer with its sources, instead of a list of links.
- Hallucination · An AI answer that sounds confident and plausible but is false or made up: a figure, a quote, a source or a reference.
- RAG (retrieval-augmented generation) · A technique that first retrieves relevant passages from your documents, then hands them to the model so its answer is based on them.
- Large language model (LLM) · A program trained on huge amounts of text to predict what comes next in a text; it is the engine behind AI assistants.
Try it
Full glossary
Frequently asked questions
How do I find a model's knowledge cutoff?
AI makers often list it in their models' technical documentation. Asking the model directly is less reliable: it can be wrong about its own cutoff.
Why does a newly released model have an older cutoff?
Between the end of data collection and public release, the model has to be trained, tested and made safe. That gap is usually measured in months.
Does web search solve the problem?
It greatly reduces it for recent facts, since the assistant reads up-to-date pages. But the model's reasoning still rests on its training knowledge, and it can misread or mis-summarize what it finds.
Does the AI know today's date?
Not necessarily. Many assistants quietly receive today's date at the start of each chat, but a model used without that setting may assume it's still around its cutoff.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.