Model routing puts a decision in front of the model call. A rule, a classifier or a small model looks at the task and sends simple ones, such as extracting a field or classifying an email, to a small, cheap model, and sends complex ones to a large reasoning model. Gartner calls this inference tiering and recommends it as the main defence against inference costs that rise as agents use more tokens per task.
The saving can be large, because prices between model tiers often differ by a factor of five or more. The risk is quality: a task routed to a model too small for it fails quietly, so routing needs its own evaluation, with a sample of each tier’s answers checked against the larger model’s.
Routing changes the rate in a monthly cost bridge, which makes its effect easy to see when cost is tracked per outcome. Without that tracking, a routing rule that saves money and a routing rule that quietly lowers quality look the same on the invoice.