what is model fallback routing?
fallback routing is when a safety classifier redirects your request to a different, usually less capable model. anthropic does notify you when it happens, but in practice that notice is easy to miss inside a long autonomous run. you keep working. the model answering you changed.
why this matters
a stricter safety classifier can decide mid-session that your request needs to be handled by a different model, often to enforce a safety or capacity limit. anthropic's own july 1 2026 redeployment of Fable 5 is a documented example: a new cybersecurity classifier improved safety coverage, at the cost of flagging more benign coding and debugging requests, and those flagged requests get rerouted to Opus 4.8 instead. the notification exists, but if you never check which model actually responded, you're trusting output from a system you didn't choose for the task.
how to catch it
log which model answered every session, and treat any fallback as a stop signal to re-scope, not a reason to just retry the same prompt. see the full breakdown at you might not be running the model you benchmarked.
source: Claude system card and anthropic's own july 1 classifier statement. last verified: july 26 2026. written by elisabeth hitz, certified in anthropic's ai fluency program (framework & foundations, and ai capabilities & limitations), plus claude 101 and claude cowork.