Claude Code effort levels and how to switch model

You can set the model in five places. When more than one is set, the higher one on this list wins:

1. In a session: /model opens the picker, /model <alias or name> switches immediately.
2. At launch: claude --model <alias or name>.
3. Environment: ANTHROPIC_MODEL=<alias or name>.
4. Settings: the model field in settings.json.
5. Default for new sessions: ANTHROPIC_DEFAULT_MODEL.

claude --model opus
claude --model sonnet --effort high

In the /model picker, Enter switches and saves the model as your default, and s switches for this session only. Left and Right adjust the effort slider for the model you are choosing. /status shows which model the session is really on.

Aliases follow the newest model in each family, so you rarely need a full model id. What each resolves to on the Anthropic API in Claude Code 2.1.309:

On Bedrock, Google Cloud and Foundry the same aliases can point at older models. Fable is not the default on any plan, and depending on your plan and seat it can bill to usage credits rather than your included limits. For plan mode itself, see Claude Code permission modes.

Effort trades tokens for depth. Lower levels answer faster and spend less; higher levels think longer and check more. The levels depend on the model:

Pick a level the model does not have and Claude Code uses the highest one below it, so xhigh runs as high on Opus 4.6. Anthropic notes that max shows diminishing returns and can overthink, so try it on the task before making it your habit.

Effort comes from the first of these that is set:

1. An explicit choice: the CLAUDE_CODE_EFFORT_LEVEL environment variable, --effort at launch, or /effort in the session.
2. Your settings: a level saved for the model, or the effortLevel key.
3. The model default from the table above.

/effort              # open the slider
/effort high         # set it directly, saved as your default
/effort auto         # clear the saved level for this model
claude --effort low  # this session only

In the /effort slider, Enter saves the level as your default and s applies it to this session only. To pin a default in a file, use effortLevel in ~/.claude/settings.json:

{
  "model": "opus",
  "effortLevel": "xhigh"
}

max is the exception people trip over. Neither effortLevel nor the pickers save it; it lasts for the current session. The one way to keep it is the environment variable, for example export CLAUDE_CODE_EFFORT_LEVEL=max in your shell profile. While it is set, it also beats a saved effortLevel and the effort a skill or subagent asks for, so unset it when you want the pickers back in charge.

• Effort is sent with every request and decides how hard the model works on the whole session.
• ultrathink anywhere in a prompt asks for deeper reasoning on that turn only. Claude Code adds an instruction to the context; the effort level sent to the API does not change.
• Thinking on or off is a separate toggle: Option+T on macOS, Alt+T on Windows and Linux, or alwaysThinkingEnabled via /config. It cannot be turned off on Opus 5.5, Sonnet 5.5, Haiku 5.5 or the Fable models.

You are billed for thinking tokens even when the thinking is collapsed or hidden.

Effort is one of the two dials you control for how quickly a subscription runs down; the model is the other. Every level above the one you need spends tokens on reasoning the task did not require, so a sensible pattern is the default for normal work, /effort low or medium for renames, tests and boilerplate, and xhigh or max for the one hard bug. /usage (also /cost) shows session cost and what is contributing to your limits. Switching to sonnet or haiku for routine work has the same effect.

/fast is not an effort level. It runs Opus 5.5, Opus 5 or Opus 4.8 up to 2.5 times faster at a higher per-token price, with the same quality. On Pro, Max, Team and Enterprise it bills to usage credits only, never to your included limits, and Team and Enterprise need an Owner to enable it. Do not turn it on to save usage; it does the opposite.

Termdeck runs the real Claude Code CLI on your machine, so the models and levels are the ones your installed version offers. The message box has a model picker and a thinking level picker for each chat, and the level list comes from the model itself, so a model without an effort dial says so and a new model arrives with its own levels. Defaults for new chats live in Settings, then General, and each chat can change its own afterwards.

Each chat keeps its own model and effort, so one chat can grind on a hard bug at a high level while another runs routine work on a cheaper model, on the same machine or different ones. The sessions docs cover the pickers, and the Codex CLI models and effort guide covers the other engine.