Knowledge cutoff

How it works

The knowledge cutoff is the point in time after which the model's training data becomes sparse or absent. Some makers publish both a reliable-knowledge date and a later training-data date; the model may know a little about the period between them.

The model has no built-in sense of today's date, so unless the date is in the prompt it may assume the world looks as it did at the cutoff. Library versions, API signatures, prices and recent events are the areas most affected.

For anything time-sensitive, put the current facts in the prompt, give the model a search or documentation tool, or use retrieval. A later cutoff helps with newer frameworks but does not replace supplying the facts.

Worked example

If a model's cutoff is, illustratively, January 2026 and a library shipped a breaking 3.0 release in April 2026, the model will likely write code for the 2.x API. Pasting the 3.0 migration notes into the prompt fixes most of this.

Common mistakes

• 1Trusting the model's answer about the latest version of a library, API or model.
• 2Not putting the current date in the system prompt, so date math and references to 'recent' events are wrong.
• 3Asking the model for its own cutoff; its self-report is often wrong, so use the maker's docs.