Billing

Before a request runs, we set aside enough credit to cover it: a generous estimate of the prompt plus max_tokens. If you don't set max_tokens, we apply a limit of 8,192 output tokens for that model and reserve for that. When the request finishes, you're charged for the tokens actually used and the rest is released. If your available balance can't cover the reservation, the request returns 402. Setting a sensible max_tokens keeps reservations small.

Trial credits are used before paid credits.

If you cancel a streaming response, or it reaches our 15-minute time limit, you pay for the prompt and the output already sent to you. Responses that break off because of a problem on our side or a partner's are not charged.