Straight to the answer
| Situation | Pick | Why |
|---|---|---|
| Complex planning, multi-step reasoning | Opus | Holds up over long chains far better than its smaller siblings |
| Everyday code, docs, conversation | Sonnet | Best value by a wide margin, the default for 90% of tasks |
| Bulk classification, extraction, filtering | Haiku | Cheap, fast, plenty for simple work |
Three things to judge on
1. Task complexity ≠ prompt length
A long prompt doesn’t mean a hard task. “Read 50 pages of contract and pull out the breach clauses” is easy (Haiku is enough). “Design a negotiation strategy from this contract” is hard (only Opus holds up).
2. Cost of being wrong
Is one mistake a retry (use Sonnet), or an actual loss (Opus, no debate)?
A bad code comment costs nothing. A bad financial report costs money.
3. Volume and latency
At high QPS, or anywhere latency matters, Opus isn’t just expensive. It’s slow. Haiku is the best pick for classification, first-pass filtering, and routing.
One combination that works
Tier it like a pyramid:
Haiku → first-pass filtering, routing, classification
↓
Sonnet → main business logic
↓
Opus → key decisions, complex planningLet the cheap model clear out the easy 80% first, and save the expensive one for the 20% that is genuinely hard. On the same budget, the overall result beats running Opus for everything.
One thing not to forget
Models keep shipping. This table is good for about three months. After every release, rerun your own eval set. Don’t do the vendors’ marketing for them. Let the numbers decide.