Choose a model policy

Your model stays yours by default

The guided setup offers Keep my chosen model, Balanced routing and Economy routing. Preserve is the default, especially for coding agents. Opting into routing explicitly approves a small reviewed set of cheaper candidates within the same provider.

Balanced and Economy use deterministic request-shape rules, not a second paid model call. Eligible simple text requests may change model. Tools, reasoning, images, signed/cache-sensitive content, stateful Responses and unknown model families retain the original model when compatibility is uncertain. Economy permits a larger simple-text envelope; it does not promise equivalent output quality.

Decide before the budget check

Routing runs before cost estimation and reservation. Leash meters the model actually used. Existing policy allowlists and virtual-key scopes still restrict both the requested model and the destination; routing cannot add access to another provider or model.

A proxy-routed request is not routed again by the gateway. No automatic retries or additional paid fallback calls are added. Routing metadata records requested model, used model and the reason without storing prompt content. Fixed-model pins remain available in CLI configuration.

Measure quality on your tasks

A cheaper eligible model is a policy decision, not proof that it is best for a task. Compare task completion and actual usage before broadening the policy. No universal multiplier or invoice saving is guaranteed.