Clawy / Guides / Which Claude model
Opus 4.8, Sonnet 4.6, or Haiku 4.5 — a practical pick-per-task guide, not a benchmark dump.
Most teams over-spend by defaulting every call to the biggest model. The three current Claude tiers span a 5× price range, and the right answer is usually "the cheapest one that clears the task." Here's how they compare and how to decide.
| Model | Price (in / out per 1M) | Context | Best at |
|---|---|---|---|
| Haiku 4.5 | $1 / $5 | 200K | fast, cheap, high-volume: classification, extraction, routing, simple chat |
| Sonnet 4.6 | $3 / $15 | 1M | the balanced workhorse: most production tasks, coding, summarization, RAG |
| Opus 4.8 | $5 / $25 | 1M | hardest reasoning, long-horizon agents, deep research, tricky debugging |
(Anthropic also offers Claude Fable 5 — its most capable model, priced above Opus — for the most demanding long-horizon work. For everyday production, the three above cover the vast majority of use cases.)
Is the task simple and high-volume (label this, extract that, route, classify)? → Haiku 4.5. It's 5× cheaper than Opus and plenty smart for these.
Is it a normal production task — coding help, summarizing, answering over retrieved docs, a chatbot? → Sonnet 4.6. Best balance of intelligence, speed, and cost.
Does it need genuine multi-step reasoning, agentic tool-use over a long horizon, or it's just hard and getting it right matters more than cost? → Opus 4.8.
Not sure? Start one tier down from your instinct and only escalate if quality misses. You'll be right more often than you expect — and cheaper.
On Sonnet 4.6 and Opus 4.8 you can dial reasoning depth with adaptive thinking and the effort parameter (low → max). A cheaper model at higher effort sometimes beats a pricier one at low effort — and costs less. Haiku doesn't support the effort parameter, which is part of why it's the speed/cost tier rather than the reasoning tier.
Because output is billed 5× higher than input, and bigger models tend to be used on bigger prompts, the real bill gap between Haiku and Opus on the same workload is often more than 5×. If you're running serious volume, the model-selection decision dwarfs almost every other cost lever.
→ See the actual numbers. Plug each model into the free cost / margin calculator with your real prompt sizes and compare cost per call side by side.Rule of thumb: Haiku for volume, Sonnet for most things, Opus for the hard 10%. Measure, then escalate deliberately.
→ Or let routing decide. Clawy's $0.04 auto tier picks a cost-optimized model per call over x402 — USDC per request, no API keys. 5-minute quickstart → → Get the AI API Pricing Kit (€19). Compare model costs in a real Excel model, plus drop-inmargin-guard code and an offline calculator — so the model you pick never costs more than you charge.