fallback

A fallback is a substitution you did not ask for. You send a request to a top-tier model, something in the system decides that request is risky, overloaded or too expensive, and a different model answers instead. You still get a reply, which is the point, but it may be noticeably less capable than the one you were expecting.

Most fallbacks are triggered by a guardrail model watching for sensitive topics, though capacity limits and cost controls cause them too. Providers rarely announce it in the moment, so the usual clue is an answer that feels oddly shallow for the model you thought you were using. If you are paying for a specific model, it is worth knowing that “which model actually answered” is not always the same as “which model you selected”.