Follow-up answer: Log every request's input/output token counts, model used, and timestamp. Aggregate in a time-series database (InfluxDB, Prometheus). Build alerts for daily spend thresholds. Segment ...