Editorial draft available to read. Factual and language review is pending. Examples are illustrative.
Decision and method
List components that generate usage: responses, retrieval, tool calls, storage and retries. Separate estimated cost from invoiced cost and define limits by session, project and period. Assign someone to review alerts and describe behavior at a limit, including user continuity and operations already running when the threshold is reached.
Illustrative example
Illustrative example: a simple question uses a short answer while document analysis performs several searches. The dashboard shows those usage categories separately. When a task exceeds its agreed budget, the assistant preserves progress and requests a decision about continuing instead of automatically launching additional steps to fulfill an open-ended promise.
Acceptance check
Simulate a long conversation, repeated failures and several concurrent sessions. Check shared limits and differences between observed usage and later billing. Accept when costs can be attributed to concrete work, alerts arrive before the defined threshold and stopping new operations is not falsely described as reversing charges already incurred by completed activity.
Key takeaways
- Control usage and limits by work performed while distinguishing estimates from actual billing.