Haiku, Sonnet, Opus: pick the cheapest model that does the job

Model families come in sizes, and the price gap is huge — often 10–30× between the small and large tiers. The skill is routing:

• Small (e.g. Haiku): classification, extraction, tagging, routine replies, summarizing short texts. High volume, low stakes — this should be most of your token spend.
• Middle (e.g. Sonnet): drafting customer-facing copy, multi-step reasoning, most day-to-day agent work. The default.
• Large (e.g. Opus): contracts, complex analysis, anything where being wrong costs real money.

The test: run 20 real examples of a task through the smaller model. If it clears your quality bar, you just cut that task's cost by 90%. Route by task, not by loyalty — mixing tiers inside one workflow ("Haiku sorts, Opus answers the hard ones") is exactly how the pros do it.