Choosing a Model

Claude Code lets you pick which Claude model handles a session — trading raw capability against speed and per-token cost.

Why it matters

The model is the single biggest lever on both quality and bill. A frontier model reasons through a gnarly refactor that a small one bungles; a small, fast model answers “what does this function do” for a fraction of the cost and latency. Picking deliberately — and using the largest for planning, a cheaper one for grunt work — is the core of cost optimization in practice.

How it works

The family spans three tiers; the exact version names advance over time, so prefer the tier alias.

Tier	Strength	Trade-off	Typical use
Opus	deepest reasoning	slowest, priciest	hard refactors, architecture, planning
Sonnet	strong + balanced	mid cost/speed	everyday coding default
Haiku	fast + cheap	weaker on hard tasks	quick edits, search, summaries

• Set the model at launch with claude --model <name> (e.g. an Opus or Sonnet identifier) or switch mid-session with /model.
• A default can be pinned via settings.json or the ANTHROPIC_MODEL env var — see model-output-configuration and environment-variables.
• Some configurations run a fast model for subagents (the-task-tool-parallelism) while the main thread uses a stronger one — a built-in cost/quality split.
• /status shows the currently active model; subscription tiers may gate which models are available.

Example

A pragmatic split on one feature:

# Plan the change with the strongest model
claude --model opus
> outline how to migrate the auth module to the new token format
 
# Execute the mechanical edits cheaper/faster
/model sonnet
> apply the plan file by file and run the tests

You paid Opus rates only for the thinking, and the bulk edit ran on a cheaper, quicker model — without leaving the session.

Pitfalls

• Defaulting to the biggest model for everything. Running Opus on trivial edits burns money and adds latency for no quality gain.
• Hard-coding a dated version string. Specific model IDs get superseded; an old pin can silently fall behind — prefer the tier alias or update it deliberately.
• Under-powering a hard task. Forcing a cheap model through a complex multi-file refactor often costs more via retries and wrong turns than one Opus pass.
• Forgetting /model persists. Switching mid-session changes every later turn’s cost until you switch back; check /status.

See also

• pricing-token-usage
• model-output-configuration
• cost-context-optimization