Usage & billing
Charges depend on the model, group and rules applicable to each request. The public catalog does not replace account-specific terms.
Common billing components
- Input: Prompts, conversation history and other content counted by the model's billing rules.
- Output: Generated content; some models also bill reasoning usage.
- Cache: Reads and writes are priced separately and should not be counted twice as input.
- Per request or image: Image tasks and other operations may use different billing units.
A simple estimate
Estimated cost = input tokens / 1,000,000 * input price
+ output tokens / 1,000,000 * output priceThis formula only applies to basic text requests without caching, tiers or additional conditions. 1M tokens does not mean one million Chinese characters.
Billing units and reservations
Credit models use actual credits multiplied by the catalog unit price. A request may temporarily reserve one credit’s value, subject to the account reservation settings. Final settlement uses actual usage and returns the difference. This is not a fixed per-request fee.
Each request is rounded to 0.000002 account balance units; very small amounts may round to zero. Usage history shows the final charge.
Checking charges
Find a request by time, key and model in usage history. Check tokens, cost and available cache details. A timeout, cancellation or incomplete response does not necessarily mean no usage occurred.
Model access follows the acceptance results. Check the catalog for prices and availability.
Long requests and streaming
A non-streaming request may disconnect with HTTP 524 if no response arrives after about 125 seconds. The upstream may still complete processing and charge for usage.
For long requests, use streaming (stream: true), although streaming does not guarantee avoiding timeouts. After a disconnect, check usage history first. Do not immediately resend while the result or charge is unconfirmed.
Need help? Use the public support contacts on our contact page.