Sonnet is the default. Use it for almost everything. Opus is the heavyweight, slower and more expensive, worth it for hard reasoning. Haiku is fast and cheap, ideal for triage, classification, and high-volume work. Most production setups use two of the three, often all three.
One of the most common questions we get the first time a team picks up the Claude API. There are three models on the menu. Which one should the workflow call?
The names are evocative but not obvious. Sonnet, Opus, Haiku. Anthropic groups them by intelligence-to-cost ratio, not by version number. Here is how we actually decide.
The three models at a glance
The price and speed gaps between them are real, but the intelligence gap is smaller than you'd expect. Sonnet is good at almost everything Opus does. Haiku is good at far more than its price suggests. The skill is matching the model to the job.
When Sonnet wins (the default answer)
Sonnet is the workhorse. The vast majority of production AI workflows should default to Sonnet unless there's a specific reason to use one of the others.
Sonnet wins for:
- Drafting, rewriting, and editing content (emails, summaries, briefs)
- Day-to-day chat use through Claude.ai and Claude Pro
- Code generation, refactoring, and review
- Most automation workflows: enrichment, triage, drafting, structured extraction
- Customer-facing chatbots and assistants
- Document analysis, RAG pipelines, knowledge retrieval
If you're not sure which model to use, pick Sonnet. You will almost certainly be right.
When Opus is worth the extra cost
Opus costs more and runs slower. The trade is real reasoning depth on hard problems. Use Opus when the answer matters more than the latency, and where the task involves multi-step thinking that a smaller model rushes through.
The use cases where we reach for Opus:
- Strategic analysis (market sizing, M&A research, board-level briefs)
- Complex coding tasks: large refactors, architectural decisions, debugging across many files
- Long-context reasoning across hundreds of pages
- Agent workflows where the model must plan, decide, and self-correct over many steps
- Edge cases that Sonnet got wrong (fallback escalation)
You'll know you need Opus when Sonnet's output keeps being almost right but missing something subtle. That's usually a reasoning gap, not a knowledge gap.
Opus is roughly 5x the per-token cost of Sonnet. Worth it for a 20-page strategic brief. Not worth it to draft 500 personalized outreach emails. The cost difference compounds fast at automation scale.
When Haiku is the right call
Haiku is the model people underuse. It's fast, cheap, and surprisingly capable for the price. The win case is volume.
Haiku is the right call when:
- You're classifying thousands of records (lead scoring, support ticket routing, content moderation)
- The task is well-defined and structured (extract these fields, return this JSON, pick one of these labels)
- Latency matters: real-time chat, autocomplete, in-app suggestions
- You're triaging or filtering before passing to a bigger model
- The use case is so high-volume that Sonnet's cost would make the whole automation uneconomic
The cost gap between Haiku and Sonnet is roughly 10x. If you're hitting the API millions of times, that's not a rounding error.
The pattern that wins: route, don't pick
The best production AI workflows don't pick one model. They use all three in a hierarchy.
A pattern we deploy often:
- Haiku looks at the incoming task and classifies it (easy, medium, hard)
- Easy and medium tasks get routed to Sonnet for the actual work
- Hard or ambiguous tasks escalate to Opus
- Output goes through a Haiku check for format validation before returning
This pattern shows up in real systems all the time. It's how you build automations that are smart where it matters and cheap where it doesn't. We cover the underlying decision in more depth in Claude API vs Claude Pro.
A model is not a product
One thing to remember as you read benchmarks. The models are constantly updated. The Sonnet you used last quarter is not the Sonnet you're calling today. Anthropic ships incremental upgrades inside the same model name (Sonnet 3.5, 3.7, 4.5, and so on).
The implication: don't over-tune your workflows to a specific model version. Build the routing logic so you can swap models when a better one ships. Most of our clients run a Sonnet workflow today that started on Opus six months ago, because Sonnet caught up.
The honest recommendation
If you're a small team picking up Claude for the first time:
- Start everything on Sonnet. Use it for daily chat, prototyping, and your first automations.
- Only reach for Opus when you have a specific reasoning-heavy task and you've verified Sonnet can't handle it.
- Reach for Haiku the moment your volume crosses a threshold where Sonnet's cost stops making sense. For most teams, that's at roughly 10,000 calls per day, not before.
If you want help mapping which of your workflows belong on which model, that's exactly what our Automation Builds engagements work on. We do the routing math so you don't have to.