Production traffic is not uniform. A support assistant handles a long tail of genuinely hard cases and a very large head of "what are your opening hours". Routing every request to the frontier model pays frontier prices for the easy ones.
A router sits in front of the model pool and decides where each request goes. Simple routers use rules: length, detected intent, whether tools are needed, which customer tier is asking. Learned routers train a small classifier on labelled traffic to predict whether the cheap model would have answered acceptably. Cascading routers try the small model first and escalate when a confidence check or a verifier rejects the answer.
The measurement that matters is not average cost. It is the quality floor on the requests that got routed down, because a router that saves sixty percent by mishandling your hardest tickets has moved cost into support.
Most teams start with a rule, log every routing decision with the outcome, and use that log to train something better once the traffic pattern is visible.