Runbook — LLM cost spike¶

Trigger: LLM cost or token rate far above baseline (dashboard / alert).

Symptoms¶

• Sudden jump in premium model usage or token volume.
• Specific agents or workloads dominating spend.

What operators do (summary)¶

1. Identify whether traffic is legitimate (launch, batch job) or abuse / bug.
2. Throttle: apply tier downgrade, concurrency limits, or per-agent quotas per internal ops guide (env / config changes not listed here).
3. Cache: confirm response cache is healthy so repeat prompts aren’t re-billed.
4. Verify spend returns to baseline; escalate product if policy change is needed.