The date after which a model saw no training data — everything later has to reach it through search, tools, or the context window.
A model past its cutoff does not know it is past its cutoff. It answers questions about newer library versions, releases, and APIs from what it saw before, in the same confident register it uses for things it knows, which is why cutoff-related errors are so easy to miss.
This is exactly the gap retrieval fills, and why documentation tools and web search matter more for fast-moving ecosystems than raw model capability does. For anything released recently, the model's opinion is a guess and the fetched page is the answer.