An endpoint calls Claude synchronously and holds the request thread for the full generation. Under morning peak the service starts timing out, though per-call latency is unchanged. What's the fix?

Answer: B — Move the call to an async path or queue and return a job reference

Flat per-call latency with degraded throughput under load is thread exhaustion, not model behaviour. No prompt or tier change fixes a blocked pool.