Model Router in plain English.

A model router is a system, sitting in front of multiple AI models, that automatically decides which model should handle each incoming request — sending simple questions to a small, cheap, fast model and reserving an expensive frontier model for requests that genuinely need it. Rather than build a product around a single model and pay frontier prices for every request regardless of difficulty, teams increasingly route dynamically: a quick classification step, or the router's own lightweight model, estimates how hard a request is and picks accordingly. Several routing products, along with in-house routers built by larger AI products, all work on this principle. The upside is substantial cost savings with little quality loss, since a large share of everyday requests don't need frontier-level reasoning. The risk is a router misjudging difficulty and sending a genuinely hard question to a model too weak to handle it well, so most production routers are tuned conservatively toward the more capable model when uncertain.