Back to full roadmap
topiccore
Cost Monitoring
Per-endpoint token spend, user × model breakdown, daily/weekly anomaly alerts.
2 hours2 resources1 prereqs
Log every LLM call: model, input/output tokens, cost (computed), user_id, endpoint, trace_id. On dashboards:
- Daily spend trend (with anomaly alerts)
- Top 10 expensive endpoints
- Per-segment user cost
- Model mix ratio
Tools: Helicone (auto), Langfuse, OpenLLMetry, custom Grafana.
Alert: daily cost > 150% normal → notify. A stuck infinite-loop agent can burn $100/hour.