Models
Knowledge cutoff
The knowledge cutoff is the date after which a model has no training data, and therefore no reliable knowledge of events.
A model asked about something after its cutoff will often answer anyway, producing plausible and wrong output rather than declining. It has no reliable sense of what it does not know.
This is one of the strongest arguments for retrieval: supplying current material in the prompt sidesteps the cutoff entirely and makes the answer checkable.
In practice
The cutoff is when training data stops, not when the model was released, and the two can be a year apart. Anything after it is absent or, worse, half-present from sources that discussed it early. For anything current, retrieval is not an optimisation — it is the only mechanism that supplies the fact.
Common questions
How do I work with information after the cutoff?
Supply it through retrieval. There is no prompting technique that gives a model knowledge it was not trained on, and asking about recent events without supplying them is a reliable way to get invention.
Is the cutoff the same as the release date?
No, and they can be a year apart. Events shortly before the cutoff are also represented unevenly — present in some sources, absent from others — which makes the boundary fuzzier than a single date suggests.