Billing & usage
Credit, token accounting, CSV exports, and refunds.
Cash and usage credit
Balances and API costs are denominated in usage value at OpenAI list rates. You buy that value at a 20% discount. $80 buys $100 of usage credit; $20 buys $25. Requests deduct the usage value, not the discounted cash amount.
How usage is measured
Billing uses provider-reported usage, not our estimate of the prompt length. Input tokens include cached and cache-write tokens. Output tokens include reasoning tokens; those subsets are not counted a second time.
| Component | Calculation |
|---|---|
| Uncached input | (Input − cached − cache-write) × input rate |
| Cached input | Cached tokens × cached-input rate |
| Cache write | Cache-write tokens × applicable write rate |
| Output | Output tokens × output rate |
Amounts are stored as integer micro-USD: 1 USD = 1,000,000 micro-USD. Each component rounds up to the nearest micro-USD. There is no floating-point balance arithmetic.
Reservations and final settlement
Before forwarding a request, the proxy reserves estimated maximum credit. Available credit equals balance minus active reservations. Output ceilings affect the reservation, so setting a sensible limit can help when your balance is low.
Once usage arrives, the final amount is deducted exactly once and the unused reservation is released in a database transaction. Requests can be refused with 402 when available credit or a spend limit is insufficient.
A response can cost more than its estimate, so a balance can temporarily become negative; further requests are blocked until credit is added. Failed upstream HTTP requests release their reservation with zero cost.
Missing usage and interrupted responses
If an upstream success has no usable usage report, the reserved estimate is charged and the event is marked usage_missing for operator review. This is an exception path, not an exact provider token count. Contact support with the request ID if you need it reviewed.
Disconnected streams are drained to recover usage where possible. A timeout or crash can prevent final observation; reconciliation and anomaly review are necessary operational safeguards.
Inspect and export usage
Choose a 7-, 30-, or 90-day period on the Usage page. View totals, daily usage, model and key breakdowns, and the latest 100 requests for that period. Exact UTC daily values are available below the chart.
The CSV export includes up to 10,000 recent rows for the selected period, scoped to your active organization. It contains metadata only: token counts, model, endpoint, masked key display, cost, status, latency, and timestamps.
Purchases, refunds, and credit validity
Owners and admins can add credit when card purchases are enabled. Payments are handled by Stripe; we never store card numbers. Credit is added only after the payment webhook confirms the amount and payment status.
Some accounts use manual funding instead. Follow the instructions on your Billing page. Unused credit is refundable on request at the cash paid for the unused portion, subject to the service terms. Credit may have a service validity deadline; contact support for your account’s applicable dates before purchasing.