Sonnet 4.6: the workhorse most candidates should default to

On the exam, when a scenario doesn't name a model, assume Sonnet 4.6. That's the production default Anthropic recommends and the model most exam questions implicitly target.

Why Sonnet 4.6 is the default

• Reasoning ≈ Opus 4.6 on most agentic benchmarks at a fraction of the cost.
• 200K context out of the box — enough for almost every agent loop with compaction.
• Tool-call reliability on par with Opus for well-designed schemas.

When to step up to Opus 4.7

• The task requires multi-step reasoning the model has to plan rather than execute.
• You need the 1M context for genuinely long inputs (codebases, document collections).
• The cost of a wrong answer dwarfs the per-token delta (high-stakes financial, legal, or code changes).

When to step down to Haiku 4.5

• High volume, low complexity (classification, extraction, formatting).
• Latency-bound user-facing paths where p99 matters.
• Sub-agent workers in an orchestrator-worker pattern.

Model ID: claude-sonnet-4-6. Worth memorizing.