There isn't one Claude — there's a family that trades intelligence against speed and cost. Opus is the deepest reasoner, Sonnet the balanced workhorse, Haiku the fast-and-cheap one. An architect routes each task to the cheapest model that still clears the bar. Pick a workload and see which fits.
| Model | Best at | Relative cost / speed |
|---|---|---|
| Haiku | high-volume, well-defined tasks: classification, simple extraction, routing, guardrail checks. | cheapest · fastest |
| Sonnet | the default workhorse: agents, coding, RAG, most production features. | mid · fast |
| Opus | the hardest reasoning: complex agents, novel architecture, high-stakes analysis. | priciest · deepest |
Route, don't pick once. A great pattern (from Domain 1's routing): use Haiku to triage and answer easy cases, and escalate only the hard ones to Sonnet or Opus. You get most of the quality at a fraction of the cost. Pair this with prompt caching and the Batch API for the biggest savings.
Intuition only sticks with numbers. Model pricing is per million tokens (MTok), input and output priced separately. Set your monthly volume and per-request sizes; the bars show what each model would cost. Then flip on the two big levers — Batch API (~50% off, for non-realtime bulk) and prompt caching (repeated prefix reads at ~10%) — and watch them stack:
Illustrative per-MTok pricing (input / output): Haiku $1 / $5 · Sonnet $3 / $15 · Opus $15 / $75. Always check the current pricing page — the ratios and the reasoning are what matter here.
Short input (~400 tokens), tiny output (~50), and a well-defined task. Run the eval: Haiku scores 94%, Sonnet 96%, Opus 96.5%. The math: Haiku ≈ $650/mo, Sonnet ≈ $1,950, Opus ≈ $9,750. Is 2 points of accuracy worth $1,300/month? For ticket routing (a mis-route just lands in a review queue) — no. Decision: Haiku, with the eval as proof. The upgrade twist: route the ~10% of tickets Haiku marks low-confidence to Sonnet — total ≈ $780/mo, near-Sonnet quality at 40% of Sonnet's price. That hybrid is the "router" pattern paying rent.
~88k requests/mo, but each carries big context (~6k tokens of repo + rules) and real output (~800). This is multi-step reasoning over code — Haiku's eval score drops hard here, so it's out. Sonnet ≈ $2,640/mo raw. Now the levers: the repo context + system prompt is a stable, repeated prefix → cache it, and input cost collapses: ≈ $1,500/mo. Opus would be ≈ $13,200 before caching — reserved for the occasional gnarly architecture question, invoked explicitly, not as the default. Decision: Sonnet + caching, Opus on-demand.
~2k tokens in, ~300 out, schema-constrained (Domain 4), and nobody is waiting on the response — it runs overnight. Haiku ≈ $525/mo in a real-time loop… but "overnight" is the tell: use the Batch API → ≈ $262/mo. Garbled docs that fail validation twice get flagged and re-run through Sonnet (~5% → +$50). Decision: Haiku + Batch + Sonnet fallback ≈ $310/mo — versus ≈ $19,500 if someone had "just used the best model" in a real-time loop. Same task, 60× cheaper, measurably equal output.
Curated companion: Anthropic — Models overview.