Not every request needs the same model. A billing classifier, a one-line rewrite, a retrieval query expansion step, and a multi-file code review have different accuracy, latency, and cost requirements. Sending all of them to the largest available model is easy, but it is rarely good engineering.

The goal is not to make everything cheap. The goal is to spend model capacity where it changes the outcome. Straightforward tasks can often run on small or mid-tier models. Ambiguous, high-risk, or multi-step tasks usually deserve a stronger model, or at least an escalation path that checks whether the cheaper answer is good enough.

Model routing is the part of the system that makes that decision. It can route by feature, query difficulty, user tier, risk level, latency budget, context length, tool requirements, or observed confidence. A good router also checks cache first, tracks quality by route, and fails closed when a wrong answer would be costly.