What is a knowledge cutoff?
Your assistant will describe last year's pricing page with total confidence, because nobody told it the page changed.
Knowledge cutoff
training cutoff
A knowledge cutoff is the date through which a language model's training data was collected. Ask it about anything after that date from memory alone, and it has nothing. No training example ever mentioned it.
Anthropic's own documentation splits this into two dates, not one. "Training data cutoff" is the full range of material used. "Reliable knowledge cutoff" is the narrower date through which that knowledge is most complete. Coverage of an event keeps growing for months after it happens, so the reliable date sits earlier than the full range. Claude Sonnet 5's training data cutoff is January 2026, and Opus 5's is May 2026.
The gap between a model finishing training and reaching a user is also real. A model can carry a January cutoff and still ship to the public months later. The release date on a product page tells you nothing about what the model actually knows.
Where this breaks a real answer
A cutoff blocks more than news. It blocks anything that changed since training stopped. Most of what a business asks about changes constantly: prices, staff, stock levels, contract terms, a competitor's latest feature. The model was never wrong about the world on the day it was trained. It is wrong now, on the things that moved since.
The failure mode is worse than a blank answer. A model with no data past its cutoff still answers fluently, because fluency and accuracy are separate properties. It describes last quarter's pricing as current. Nothing in the response signals that the information is stale.
This is the business case for retrieval, stated plainly: your data moves every day, and model weights do not move at all between training runs. A support bot that answers from the model's memory will eventually quote a return policy you retired months ago.
- TrainCutoff date locked in.
- ShipRelease, months later.
- AskQuestion about today.
- RetrieveFetch current source.
- AnswerGrounded, not guessed.
Retraining resets the cutoff for everything at once. Retrieval fixes one question at a time, using data that is current the moment it is fetched.
Related questions
01Does a bigger context window fix a knowledge cutoff?
No. A context window only holds what you put into it for that request; it does nothing to what the model learned during training. A larger window lets you paste more current documents into the prompt, but something still has to fetch and choose those documents. That fetching step is retrieval, not the window itself.
02Why does ChatGPT sometimes know about recent events?
Because it is not answering from memory. Browsing-enabled products run a live web search and feed the results into the prompt, then write the answer from that fetched text. The underlying model's training cutoff has not moved; only the input it was given has.
03How do I know a model's exact cutoff date?
Check the vendor's own model documentation rather than asking the model, which does not reliably know its own cutoff and often gets it wrong. Anthropic publishes both a training data cutoff and a narrower reliable knowledge cutoff for each Claude model, since the two dates answer different questions.
