Classification, extraction with clear schemas, routing decisions, and formatting — high-volume tasks with definable correctness, where latency and unit cost dominate.
Complex reasoning, long-document synthesis, high-stakes agent decisions, and anything where a quality miss is expensive — the top of your routing hierarchy, used deliberately.
Classify difficulty → dispatch to the cheapest adequate tier → escalate on low confidence. Your evaluation set defines 'adequate' per task; re-benchmark as releases land.
Data-boundary mandates and extreme volume can justify self-hosted small models — priced honestly against ops burden, not ideology in either direction.
Skipping the discipline this article describes until an incident, audit, or stalled project forces it — every practice above is cheaper adopted early than retrofitted under pressure.
Let's discuss how we can help you with small vs frontier models.