You should never pay for a failed generation

Generative models fail sometimes — a provider hiccup, a moderation block, a timeout. You shouldn't eat the cost when there's no output to show for it.

How refunds work

• Credits are debited when a task is created.
• If the upstream task fails before producing a result, the debit is reversed automatically.
• Refunds are idempotent — a retried failure never double-refunds.

The result: your effective spend tracks successful output, not attempts.