Small models, big jobs: the routing economics most teams ignore

Watch a typical production AI system and you'll find the same premium model handling everything: classifying a support ticket (trivial), extracting a date from an email (trivial), and drafting a nuanced client proposal (genuinely hard). Two of those three calls are overpaying by an order of magnitude.

The routing principle

Classification, extraction, formatting, and simple triage are small-model jobs — today's small models handle them at accuracy indistinguishable from frontier models, at a small fraction of the cost and latency. Reasoning, long-form writing, judgment calls, and anything customer-visible with nuance stay on capable models. The router can be a rule ('all triage → small model') or a cheap model making the call ('is this simple or complex?').

The numbers from member projects are consistent: 60-80% cost reduction at equal measured quality, with latency improvements as a side effect — small models are also fast models, which matters enormously for voice applications.

The one non-negotiable

Routing without evaluation is guessing. The only reason members can downgrade a call confidently is that their test suite proves quality held. Build the eval first (Professional track, Module 11), then route. If you route first, you'll discover quality problems the way you least want to — through users.