Route Before You Answer

Do not send every request to the same model. Classify first, then route.

Implementation:
1. A cheap model classifies each request into cheap, standard or frontier, returning strict JSON with a confidence score.
2. Below 0.6 confidence, escalate one tier. A wasted expensive call is cheaper than a wrong cheap one.
3. Log the tier distribution. In most production workloads 70 to 90 percent of traffic lands in cheap, and if yours does not, the classifier is miscalibrated.
4. Sample the cheap tier for quality regularly. Routing failures show up as quality drift, not as errors.
5. Keep the classifier prompt short and cached. It runs on every request, so its own cost matters.

Typical result is 60 to 80 percent lower spend with no measurable quality change on the routed-down requests.