Knowledge Cutoff

A knowledge cutoff is the date after which a model has no training data — it simply doesn't know about events, product releases, API changes, or prices that happened later. Ask a model with an early-2024 cutoff about a framework released in 2025 and it will either say it doesn't know or, worse, confidently invent an answer. The cutoff is a property of the pretraining data, not the model's release date, and the two can differ by months. For SaaS builders this matters constantly: any feature that touches current facts — news, live prices, your own changing docs, today's inventory — cannot rely on the model's parametric memory. The fix is retrieval-augmented generation or tool/web-search calls that inject fresh, authoritative data into the prompt at request time. Always tell users when an answer may be stale, and never hard-code assumptions about what the model "knows" — cutoffs move forward with every model generation, so a limitation you designed around last year may no longer apply.