Three tiers, three jobs
Claude Code lets you swap the underlying model per session, per skill, or per subagent. The three families currently available — Haiku, Sonnet, Opus — trade speed, capability, and cost. Knowing which to reach for is half the leverage.
Haiku. Fastest, cheapest. Right for high-volume read-heavy work — exploring an unfamiliar repo, running a fleet of background agents, automating dozens of CI tasks. Smart enough for "find / count / categorize" type prompts; not your first pick for architecture decisions.
Sonnet. The everyday default. Strong coding, strong reasoning, sane cost. If you're not sure which to pick, this is the answer. Most live coding sessions live here.
Opus. Slowest, most capable, most expensive. The model you switch to when stakes are high — security audits, architecture design, gnarly multi-file refactors that need long-range thinking. Combine with /effort high|max for max reasoning.