Model pricing
Model prices and billing units.
Image generation
Images Generations is billed by the number of images actually returned successfully. n is the maximum number of images attempted by the request and currently supports 1-4. If a request uses n=4 but returns only one image, only one image is charged.
An invalid n fails before any upstream request is sent, so it creates no usage log and no charge.
Text models
Token-priced model products can use up to four price components. The applicable components depend on the model product published for the current API entry:
- Ordinary input: input tokens not counted as cache reads or cache creation.
- Cache-hit/read input: input tokens served from an existing cache.
- Cache-creation/write input: input tokens written to a cache when the API/wire family reports that usage; currently this applies to Anthropic Messages.
- Output: tokens generated by the model.
Prices shown in the console are authoritative for the current account; a same-named public model id on another API entry may map to a different product.
Real upstream usage takes priority. For token-priced products, a valid non-stream success JSON object with missing or invalid usage may use an output estimate with a one-output-token minimum, even with no extractable output (including {}). Stream estimates require extractable generated output; heartbeats, empty events, and raw chunks are insufficient. An upstream disconnect, client cancellation, or delivery failure may still settle trusted partial usage or extractable output already observed by the gateway; successful chunk delivery is not required.
Available balance must be positive before a call
Account balance minus active request reservations must remain greater than zero before a model call can start. Settlement may take the account balance below zero, but an available balance of zero or less blocks another model call.
Last updated on