Tokens are charged once generation succeeds. Tokens are not returned when the result is unsuitable because of the prompt or model behavior; they are refunded automatically after a technical service error.
This page in other languages