Models

Only models you switch on are forwarded. Anything else is 400 model_not_allowed. A model Fuse cannot price is 400 model_not_priced. Built-in prices come from the catalog. A custom model needs the input and output price you enter, in USD per 1M tokens.

When a price row has a long-context tier, the higher rate is reserved once the input crosses that threshold.

The allowlist is the list of model ids on the project, not a prefix. gpt-4.1 does not allow gpt-4.1-mini. Turning a model on without a key for its provider still fails later with provider_key_missing. The dashboard locks those switches until the key is saved.

Custom models are priced by you. Enter input and output as USD per 1M tokens. Cached input is optional. The id must be exactly the string the provider expects. Removing a custom model drops it from the allowlist.

Price changes in the catalog apply to new requests within about a minute. A request already reserved keeps the estimate it took.