Overnight, the daily cost of a multi-step Claude pipeline rises sharply with no change in request volume. You are writing the runbook entry for this symptom. What should the first diagnostic step be?

Answer: A — Check per-step token usage and cache read rates to find which step's input or output grew

(a) is right: usage data localizes the increase, revealing a cache that stopped hitting, a retry loop, a bloated tool result or a prompt change, and it costs nothing to check. (b) looks outside before checking your own telemetry. (c) is a quality-risking reaction before diagnosis. (d) removes a cost saving and can make the spike worse.