Model Selection

Claude Code lets you pick the model per session. As of June 2026 the lineup is Opus (most capable, deepest reasoning), Sonnet (balanced: strong and noticeably cheaper), and Haiku (fastest and cheapest, for narrow mechanical work). Current versions: Opus 4.8 (2026-05-28), Sonnet 4.6 (Feb 2026), Haiku 4.5 (Oct 2025); the newest family is Claude 4.X. Opus 4.8 defaults to high effort (/effort xhigh).

The practical rule: match the model to the hardest step, not the whole task. Reach for Opus on architecture, gnarly debugging, and multi-file reasoning; drop to Sonnet for routine implementation and most day-to-day work; use Haiku for tight, well-specified edits and high-volume fan-out where each unit is simple.

How you use it day-to-day

• Switch with /model. It changes the current session only. Press d in the picker to set the default for new sessions. Change tiers as the work changes instead of committing up front.
• Fast mode (/fast) keeps Opus but speeds up output. It does not downgrade to a smaller model. On Opus 4.8 it runs at roughly 2× the standard token rate for ~2.5× the speed. Use it when you want Opus quality with less waiting.
• Set fallback models. The fallbackModel setting (and --fallback-model) takes up to three models tried in order when your primary is overloaded or unavailable, so a busy Opus doesn't stall the session.
• Delegate cheap, decide expensive. When you fan out subagents for independent grunt work, run them on Sonnet/Haiku and keep the orchestrating session on Opus.

Gotchas

• Bigger isn't always better. Opus on trivial mechanical edits mostly burns tokens and time. Right-size to the step.
• Fast mode trades throughput, not quality (same model). There's little reason not to use it on Opus when it's available.
• Versions move fast. Treat the lineup above as a June-2026 snapshot; re-check the models overview and changelog rather than hard-coding a version.