Skip to main content
Usage telemetry and safety metrics feeding an organization billing ledger
Open Usage in the console after making requests. Review spend, remaining balance, token counts, latency, model mix, prompt-cache hits, and safety outcomes. Each request can include more than one billable component: upstream model inference, SERV reasoning generation, prompt-guard evaluation, audit or repair work, and shadow-agent validation. A cache hit avoids a new SERV reasoning-generation call, but the upstream model still runs and is billed. Change one variable at a time: model, reasoning effort, prompt, tools, schema, or safety feature. Keep an evaluation set and compare task success, failure rate, latency, and total cost. Open Billing from the account menu to manage credits and auto top-up. See Models for model pricing. If your balance cannot cover the request’s estimated ceiling, SERV may reject it before billable inference begins. Use a reasonable output-token ceiling and monitor finish_reason or the endpoint’s stop reason for truncation.