Knowledge cutoff
In shortThe knowledge cutoff is the date after which a model's training data stops, so it knows nothing about later events unless you give it fresh information.
In plain words
A model learns from a fixed collection of data gathered up to some date. Whatever happened after that date is not in its training. That date is the knowledge cutoff. Anthropic's model overview shows the idea well. It lists two dates per model: the reliable knowledge cutoff, the date through which the model's knowledge is most extensive and reliable, and the training data cutoff, the broader range of data used. Knowledge near the end of the range is usually thinner than knowledge from earlier years.
Models are released and used for a long time after their cutoff. So at any moment, a model you use may be months or years behind. The model page of each provider states the date for each model.
Why it matters
- Outdated facts: Prices, laws, software versions, sports results, office holders and news can have changed.
- Code: A model may suggest functions that were renamed or removed in a newer library version.
- Wrong "today": A model does not know today's date unless the app tells it. Dates in answers can be off.
- False confidence: It will often answer as if its knowledge were current. That can lead to a hallucination about recent events.
- No self-check: Do not rely on a model to report its own cutoff. Read the provider's model page.
What helps:
- Paste the current text, such as the new documentation or the latest article, into your prompt.
- Use RAG so the model works from fresh documents. Anthropic's glossary notes that this lets a model use information beyond its training data.
- Use a search or other tool (tool use) when the app offers it.
- Say today's date in your prompt when it matters.
Example
You ask for the correct call to a software library. The model gives an answer based on the version it saw in training. Your project uses a newer version where the function has a different name, so the code fails. You paste the section of the new documentation into the chat and ask again. The answer now matches your version.
Common misconceptions
- "A web search changes the model's cutoff." It does not. It adds fresh text to the prompt for this one answer.
- "The cutoff is the release date." The release comes later, often by months.
- "Before the cutoff, the model knows everything." Coverage is uneven, and rare topics can be missing.
Questions and answers
What is a knowledge cutoff?
The date where a model's training data ends. The model does not know about events after it unless you provide the information.
How can I find a model's knowledge cutoff?
Look at the provider's model page. Anthropic, for example, lists a reliable knowledge cutoff and a training data cutoff for each model.
How do I get current answers from an AI?
Give the model current text, use retrieval or a search tool, and state today's date. Then check key facts.
Read next
- Hallucination
A hallucination is a fluent, confident AI answer that is false, invented or not backed by the sources it was given.
- RAG (retrieval-augmented generation)
RAG (retrieval-augmented generation) first looks up relevant passages in your documents, then gives them to the model so it answers from them.
- Tool use (function calling)
Tool use, also called function calling, lets a model request an outside action, such as a search or a calculation, which an app runs and feeds back.
- Large language model (LLM)
A large language model (LLM) is an AI system trained on huge amounts of text to predict the next token, which lets it write, summarise and answer.
- Context window
The context window is the maximum amount of text, counted in tokens, that a model can read and write in one request.