Knowledge cutoff
A knowledge cutoff is the date after which a model has little or no information, because its training data was collected up to around then. Events after it are unknown unless supplied in the prompt.
Also called: training cutoff, training data cutoff, cutoff date
By Divya Sharma · Checked Nov 15, 2026 · How we review
A model learns facts during pretraining, which uses data gathered up to some point. Model makers usually publish that date. Ask about something that happened later and the model either says it does not know or, worse, guesses, which is a common source of hallucinations.
Why the date is fuzzy
The published cutoff is a rough guide. A 2024 study, "Dated Data", found that a model's effective cutoff often differs from the stated one, and can differ by topic. One reason it gives is that new web crawls contain plenty of old pages. In practice a model may know the last stretch before its cutoff less well than earlier periods.
Models also often do not know their own cutoff or today's date. If the date matters, put it in the system prompt.
Working around it
The cutoff is fixed for a given model version. To give a model newer facts, pass them in at request time: retrieval-augmented generation from your own documents, a web search tool, or plain text pasted into the prompt. Fine-tuning is a weak way to add fresh facts. For coding, this is why a model may suggest an older version of a library's API; giving it the current docs fixes much of that.
Cutoff vs release date
A model is usually released months after its cutoff, and then used for a long time after that. The gap between what the model knows and what is true today keeps growing for as long as it is in use. Our model reviews list cutoffs where makers publish them.
Related terms
- Pretraining
- Pretraining is the first and largest stage of training a model, where it learns general patterns from a very large unlabelled dataset, usually by predicting the next token, before any task-specific tuning.
- Retrieval-augmented generation
- Retrieval-augmented generation (RAG) is a pattern where a system first searches a collection of documents for passages relevant to a question, then gives those passages to a language model so its answer draws on them.
- Hallucination (AI)
- A hallucination is output from an AI model that sounds fluent and confident but is false, unsupported by its sources or made up, such as an invented citation, quote, statistic or API function.
- Grounding
- Grounding is tying a language model's answer to specific, checkable sources, such as search results, documents or a database, so that its claims rest on supplied evidence rather than only on training data.
- System prompt
- A system prompt is the block of instructions an application sends to a language model before the user's messages. It sets the model's role, rules, tone and output format for the whole conversation.