Model Routing

Pick the right model for each request

Sending every request to your strongest model is simple, and it means paying reasoning prices to answer “what are your opening hours”. Sending everything to the cheapest is cheap and frequently wrong.

A router inspects each request and picks a model to match: a small fast one for lookups, a reasoning model for analysis, a vision model for images.

The trap is treating the router as infrastructure rather than as a component that can be wrong. It makes a judgment on every request, it will misroute some of them, and a misrouted hard question produces a cheap confident bad answer. Test its decisions, define a fallback, and track what it sends where.