Choosing between Opus 4.7, Sonnet 4.6, and Haiku 4.5

The current production lineup as of May 2026:

• claude-opus-4-7 — hardest reasoning, 1M context optional
• claude-sonnet-4-6 — production default
• claude-haiku-4-5-20251001 — fast, cheap, narrow tasks

The decision flow

Start with Sonnet 4.6. That's the default unless you have a reason to deviate.

Step up to Opus 4.7 if any are true:

• The task requires planning across more than a few logical steps
• The input doesn't fit 200K and you genuinely need the 1M window
• A wrong answer is expensive (auto-trading, legal drafts, large code rewrites)
• You're driving an evaluator-optimizer loop and want the strongest judge

Step down to Haiku 4.5 if any are true:

• The task is a narrow classifier, extractor, or formatter
• You're a sub-agent worker dispatched by a Sonnet/Opus orchestrator
• Latency or cost is the binding constraint and quality is "good enough" at Haiku

Mix tiers in one agent. It's normal — and often optimal — for the orchestrator to be Opus and the workers to be Haiku. The exam likes scenario questions where the right answer is *exactly* this mix.

What about Opus 4.6 and Sonnet 4.5?

Still supported, still fine in existing deploys. On the exam, when in doubt, the latest version is the correct answer unless the scenario constrains you (cost, latency, existing pipeline).