When to use Opus vs Sonnet vs Haiku ⚡
Master this high-ROI AI workflow: The mistake most people make is pouring Opus for every cup. This guide shows you when to use Opus vs Sonnet vs Haiku.
⚡ TL;DR
- The mistake most people make is pouring Opus for every cup.
- This guide shows you when to use Opus vs Sonnet vs Haiku.
The mistake most people make: pouring Opus for every cup. This guide fixes that.
THREE QUESTIONS
The decision tree
Three questions. Ask in order. Stop at the first "yes."
1. Is this strategic?
"Will the answer shape a real decision: pricing, positioning, hiring, money, a public commitment?"
- ✅ Yes → Use Opus. The stakes justify premium reasoning.
- ⬇️ No → Continue to Q2.
2. Does the wording matter?
"Will a human read this exact output? Client copy, published content, an important email?"
- ✅ Yes → Use Sonnet. The everyday workhorse, strong at language, drafting, judgment.
- ⬇️ No → Continue to Q3.
3. Is speed or volume the priority?
"Do I want this to feel instant? Is the task mechanical: formatting, extracting, tagging?"
- ✅ Yes → Use Haiku. Fast, cheap, accurate enough for routine work.
- ⬇️ No → Default to Sonnet.
💡 The 80/20 rule: When in doubt, default to Sonnet. It's the right cup roughly 80% of the time. Upgrade to Opus only when you can name the strategic stake. Drop to Haiku when you notice yourself doing the same routine task more than twice.
WORKED EXAMPLES
12 worked examples
The right model for each, plus a one-line "why."
| # | Task | Pick | Why |
|---|---|---|---|
| 1 | Drafting an Instagram caption | Sonnet | Voice and word choice matter; Haiku reads flat. |
| 2 | Outlining a 6-month content strategy | Opus | Multi-variable, strategic, hard to redo if wrong. |
| 3 | Reformatting a list into clean bullet points | Haiku | Mechanical. Opus here = waste of money. |
| 4 | Summarising 10 articles into one digest | Sonnet | Synthesis needs nuance, but not deep reasoning. |
| 5 | Competitive teardown of 3 brands' positioning | Opus | Strategic call you'll act on. Worth the cents. |
| 6 | Quick fact lookup ("when did X launch?") | Haiku | One-shot, factual, no judgment needed. |
| 7 | Reviewing a contract clause for risk | Opus | High-stakes, ambiguity-heavy, real consequences. |
| 8 | Tagging customer feedback by theme | Haiku | Repetitive classification, Haiku nails it at scale. |
| 9 | Architecting a new automation system | Opus | Design decisions compound; hard to undo later. |
| 10 | Writing a small utility function | Sonnet | Common code, but you want it correct first try. |
| 11 | Daily email inbox triage / summary | Haiku | Volume + speed beat depth. |
| 12 | Brainstorming product names | Sonnet | Creative range matters; Haiku is too literal. |
THE REAL COST
What this actually costs
Current API prices (per 1M tokens, April 2026)
| Model | Input | Output |
|---|---|---|
| Haiku 4.5 | $1 | $5 |
| Sonnet 4.6 | $3 | $15 |
| Opus 4.7 | $5 | $25 |
Tokens ≈ words ÷ 0.75. So 1 million tokens ≈ 750,000 words ≈ a small novel.
Plain-English calculator
Assume "1 hour of active use" ≈ 30k input + 15k output tokens (typical chat with some context loaded).
| Daily use | Haiku | Sonnet | Opus |
|---|---|---|---|
| 15 min/day | ~$0.80/mo | ~$2.40/mo | ~$3.95/mo |
| 1 hour/day | ~$3.15/mo | ~$9.45/mo | ~$15.75/mo |
| 3 hours/day | ~$9.45/mo | ~$28.35/mo | ~$47.25/mo |
| Heavy (5+ hr/day, big docs) | ~$25/mo | ~$75/mo | ~$125/mo |
What this means
🥛 Sonnet ≈ a Netflix subscription for daily use. Most people massively overestimate the cost.
☕ Opus is only ~1.7× Sonnet not "10× expensive" like older Opus versions used to be. The pricing repriced down in 2025.
⚡ Haiku is so cheap that pinching pennies on Haiku tasks isn't worth your time. Spend the savings on better prompts instead.
📌 Note: These are API costs (what you pay when building automations or running Claude through tools like Claude Code). If you use Claude.ai chat, you're already on the flat $20/mo Pro plan and these per-token numbers don't apply directly.
STOP OVERSPENDING
Tasks that almost always belong in Haiku
This is where most people overspend. Move these off Opus immediately.
- ✅ Summarising emails or Slack threads
- ✅ Reformatting lists, bullet points, JSON, CSVs
- ✅ Tagging or categorising (sentiment, topic, priority)
- ✅ Translating short snippets between languages
- ✅ Generating subject lines, alt text, meta descriptions
- ✅ Extracting structured data from a paragraph (names, dates, prices)
- ✅ Spelling, grammar, capitalisation fixes
- ✅ Quick definitions and fact lookups
- ✅ Title-case / sentence-case conversions
- ✅ Drafting routine messages (booking confirmations, RSVPs, intros)
- ✅ Estimating reading time or word counts
- ✅ Generating placeholder copy (lorem-ipsum-style filler)
- ✅ First-pass classification before a human reviewer
🎯 Rule of thumb: If the task has one right answer and you wouldn't pay a human more than $5 to do it, it's a Haiku task.
PRINT THIS
Wallet card
Print this page. Cut along the box. Fold once. Keep at your desk.
Verify current pricing at platform.claude.com/docs/en/about-claude/pricing.