Model selection (Opus vs Sonnet vs Haiku)

Claude ships in a family of tiers that trade capability against speed and cost. Opus is the most capable tier for deep reasoning, long multi-step agentic work, and hard analysis. Sonnet is the balanced tier — strong reasoning at lower latency and cost, the usual default for production. Haiku is the fastest and cheapest tier, built for high-volume, latency-sensitive, or simpler tasks like classification, extraction, and routing.

Selection is an engineering decision, not a status ranking. The right move is to profile the task: a simple label or a short extraction rarely needs Opus, while a nuanced multi-file refactor or a long agent loop may justify it. A common production pattern is routing — a cheap model triages or classifies, and only the hard cases escalate to a larger tier.

Model names are versioned (a tier plus a generation/date), so "Sonnet" alone is ambiguous — always pin the specific version string in code. Never assume a fixed price or context window from memory; look up the current model card, since these change across releases.