Picking a Model by Cost and Latency Tier, Not Just Capability

The biggest, most capable model is rarely the right default for every call in a system — it's the right choice for the calls that actually need it. Structure your model usage in tiers instead: Small/fast tier: classification, routing, extraction, simple formatting, anything high-volume and low-ambiguity. This is where most of your call volume should live. Mid tier: general chat, drafting, summarization — good default for anything user-facing that needs decent quality without reasoning-heavy depth. Large/frontier tier: complex reasoning, high-stakes outputs, anything where a wrong answer is expensive to have shipped. Reserve this tier deliberately rather than defaulting to it everywhere. A simple router (even a small model deciding "does this need the big model") often pays for itself quickly at any real volume. Measure cost per successful outcome, not cost per call — a cheap model that fails and gets retried by an expensive one isn't actually cheap.