Knowledge cutoff

A model learns from data collected up to some point, then is trained and released, often months later, and stays in use long after that. Anything that happened after its cutoff is invisible to it. It may not know about recent product releases, current prices, new laws, or the latest version of a library.

The trap for builders is that a model rarely says it doesn't know. Asked about something recent, it may answer from older information as if it were current, or invent a plausible update. This affects anything time-sensitive: pricing, policies, API details, news, and even the current date, which the model can't know unless you put it in the prompt.

The fix is to stop relying on the model's memory for facts that change. Pass today's date in the prompt when timing matters. Retrieve current information from your own database, documents, or a search tool. Treat the model as a reasoning and writing engine over data you supply, not as the source of truth.

Your own data is always past the cutoff, because the model never saw it. That's the core reason retrieval and tool use exist in almost every serious AI feature.