Zero Completion Insurance

Every request is charged the provider's real bill for it, converted to DA. When a model's answer comes back empty or ends in an error, that bill carries no charge for the model's tokens — so neither does your wallet. It applies to every account, every model and every provider, with nothing to switch on.

What is covered#

No charge is taken for the model's prompt, completion or reasoning tokens when:

• the answer has zero completion tokens and a blank or null finish reason, or
• the answer has an error finish reason (finish_reason: "error").

The same holds for other kinds of output: a speech answer is not billed until it delivers usable audio, and a picture answer that returns no decodable picture is not billed.

What is still charged#

• Work a helper already did. A web search, a page read or a PDF parsed before the answer failed may still be billed, on the helper's own meter. When it is, it shows in the request's usage.
• An answer that stopped early with content. A stream that ran out at max_tokens, or that you closed, delivered usable output: it is billed on what the provider produced.
• Your own error. A request refused before it runs — 400, 401, 402 — never reaches a model and charges nothing either.

How the charge is taken#

Nothing is held or reserved before a request runs. Your wallet admits the request while your balance is above zero; when the answer is done, its real bill is converted to DA and taken off your balance once. An empty or failed answer has a real bill of zero for its tokens, so nothing is taken for them.

Each request, and what it was charged, appears in your activity within seconds of the answer completing.

Note. The insurance does not replace retry logic. For production work, back off and retry on 503 and 504, and consider model fallbacks so a request can move to another model when the first is unavailable.