Models, effort & thinking

Claude Code lets you choose the model, how hard it reasons (effort), and whether it thinks step by step before answering. This page covers the controls and how to pick sensibly.

Model aliases

Alias	Resolves to
default	Your plan's default model
best	The strongest available (= opus)
opus	Opus — on the Anthropic API, Opus 4.8 (claude-opus-4-8)
sonnet	Sonnet — on the Anthropic API, Sonnet 4.6
haiku	Fastest, lightest model
sonnet[1m] / opus[1m]	1M-token context variants
opusplan	Opus while planning, then Sonnet to execute

NOTE

Model IDs depend on the provider. On Bedrock or Vertex, an alias may resolve to a 4.5/4.6 build. State the provider when quoting exact IDs.

Defaults by plan

Plan	Default model
Max / Team Premium / Enterprise	Opus 4.8
Pro / Team Standard	Sonnet 4.6

Switching models

Use /model to open the picker (it saves your default since v2.1.153), or Option/Alt+P to switch quickly. At launch, pass --model <alias>. Environment variables also apply: ANTHROPIC_MODEL sets the main model and CLAUDE_CODE_SUBAGENT_MODEL sets the model for subagents.

Effort

Effort controls how much reasoning the model spends per turn — how thoroughly it deliberates before and while it acts. It does not change the model's raw capability (that's your model choice); it dials how hard that model thinks. More effort means more careful, thorough work, at the cost of more time and tokens.

Level	How it works	Best for	Speed / cost
low	Answers fast, minimal deliberation	Mechanical edits — typos, renames, log lines	Fastest, cheapest
medium	Balanced, reasonable reasoning	Routine, well-scoped tasks	Balanced
high	Reads carefully, checks details, weighs trade-offs (default on Opus 4.8 / Sonnet 4.6)	Everyday real work	Slower, more tokens
xhigh	Explores alternatives, anticipates edge cases	Hard problems worth the extra compute	Slow, expensive
max	Maximum deliberation — considers rare cases, double-checks	The hardest work: subtle bugs, large refactors, architecture	Slowest, most expensive

Set it with /effort, the --effort flag, or the effortLevel setting. Every level above the default burns more reasoning tokens, so match the effort to the difficulty: high effort on a one-line change is wasted time and weekly usage.

TIP

Effort is the dial; the model is the engine. A stronger model (Opus) reasons better at any effort; more effort makes whichever model you're on reason harder. Need a one-off deep pass without changing the setting? Add ultrathink to your prompt (see Thinking).

Thinking

Extended thinking lets the model reason before responding. Toggle it with Option/Alt+T. For a one-off deep pass, include the keyword ultrathink in your prompt — it requests the deepest reasoning for that single message.

opusplan

opusplan is a hybrid: Claude uses Opus while in plan mode to design a solution, then switches to Sonnet to execute it. This gives you strong planning without paying Opus rates for every implementation step.

Picking a model by task

Task	Suggestion
Quick fixes, formatting, simple edits	sonnet or haiku
Everyday feature work	sonnet
Hard architecture, tricky debugging, planning	opus / best
Plan thoroughly, then implement cheaply	opusplan
Very large codebase context	opus[1m] / sonnet[1m]

TIP

For balancing capability against spend, see cost optimization.

Source: Model configuration