The date after which a model has no built-in knowledge — anything more recent must be supplied via tools, search or your prompt.
A model's knowledge cutoff is the last date covered by its training data. Ask about an event after that date and a model will either say it doesn't know or, worse, confidently invent an answer. Cutoffs typically lag the release date by several months to a year.
The cutoff only bounds built-in memory, not what a model can do. Give it web search, function calling or RAG and it works with live information regardless of when it was trained. For anything time-sensitive, design for retrieval rather than trusting recall.
Each model page on modelgrep lists the knowledge cutoff where the maker discloses it, alongside the release date — the gap between them is a useful tell about how current a model's baked-in knowledge is.