Usage and allowances
Understand consumption, reservations, and exceptions.
Find your consumption
Open Usage & reconciliation to inspect the inference ledger for your organization. Open Orders & agreements for an agreement’s total, remaining, and reserved value.
Usage records connect a request to the agreement and accepted rate card used for delivery. Public catalogue prices can change without changing historical charges.
Browse older records
Use Older records below the ledger to load the next server page and Newer records to return toward the latest entries. The displayed range describes the current page, not your lifetime request count.
If you opened usage from an agreement’s View consumption action, that agreement filter stays in place while you change pages. Audit history uses the same newer/older navigation for authorized event records.
Three amounts to understand
| Amount | Meaning |
|---|---|
| Total agreed value | The value represented by the accepted quantity and rate card |
| Remaining value | Value not yet settled as consumed |
| Reserved value | Allowance held for admitted work or unresolved reconciliation |
Reserved value is not available for another request or model amendment while the associated work is pending. Review the recorded outcome if a reservation remains unresolved.
Read actual and equivalent usage
Actual usage records input, output, and supported cache classes. Equivalent consumption expresses the same charge relative to your agreement’s input base rate.
A one-million-output-token workload can consume more than one million input-equivalent tokens when the output price is higher. See the worked example in Understanding token prices.
Cached tokens must not also be counted as ordinary input. Reasoning tokens may already be included in a supplier’s output counter; they must not be charged twice.
Follow provider allocations
For an agreement supplied by more than one provider, open Provider allocation rate cards in Orders & agreements. Each card shows its own allowance, remaining value, reserved value, delivery dates, and accepted rates.
Consumption follows the allocation used for the request. A request is served by one eligible provider allocation; the service does not combine unrelated suppliers’ output into a single model answer. Compare the request’s pricing snapshot with that allocation’s rate card when reviewing usage.
The total agreement view brings these commitments together, but one headline rate should not be applied to every provider. Previously consumed usage continues to use its original price snapshot after an approved future amendment.
Pending and exception states
A supplier timeout or missing final usage record does not prove that no work happened. If the result or metering is uncertain, the service retains the reservation and records a reconciliation exception.
Do not assume that an interrupted response is free or immediately retry a large uncertain request. Record the request and agreement references and contact your operator to review the exception.
The same rule applies to a browser-sent request: after a cancelled or interrupted Send an inference request submission, check the ledger before retrying. See Sending an inference request.
Request an exception review
Contact the team with the request and agreement references and the issue you see. A timeout or missing response does not prove that no work occurred. Review the recorded outcome before retrying an uncertain request.
Why consumption can stop
The gateway can reject a request because the agreement has not started, has ended, lacks available allowance, exceeds throughput or concurrency, requests a different model, or has no configured eligible supplier route.
An allowance is necessary but not sufficient for admission. Review the agreement window, request capability, and delivery constraints as well as the balance.
When to ask for help
If a ledger entry is unexpected, provide its request reference, agreement reference, approximate time, and the issue you see. Do not include API keys, supplier credentials, or sensitive prompt content in a support enquiry.