Documentation menu

concepts

Model routing

Per-shot, policy-driven model selection with automatic failover. Models are data, not product.

Last updated 2026-07-10

How routing works

A routing policy maps shot intent to models: dialogue-heavy shots route to a dialogue-strong model, reference-driven action to a multimodal-reference model, image-to-video work to an i2v-strong model. The policy and the model registry are reviewable configuration.

Call list_models to see the current roster with status, capabilities, and cost tier: see the API reference.

Overrides

You can override the routed model per request with model_override. The full model set is reachable from every surface, including agents.

Failover

If a model fails mid-sequence, the router excludes it, re-walks the policy, re-encodes the shot for the next model, and retries, with every hop audit-logged. Only all models down produces a failure, and it is an honest one.

Why config

Model names and prices change every few months. Keeping them as data means a model swap is a data change, not a release, and the site never claims a model that the registry does not.

Model routing | Cadenza Docs