Claude Code effort per model: set it once and never touch it again

TL;DR One level per model. Run /effort on each model you use and confirm with Enter: it lands in modelSettings.<model>.effortLevel, and that entry beats the global effortLevel in the same file. s keeps it to this session, and /effort auto clears the active model's.

If you use more than one model, you have been working two dials. First /model, then /effort, because the reasoning level you want on the model that investigates for an hour is not the one you want on the model renaming three variables. Since v2.1.251 the second dial travels with the first: Claude Code saves the level per model, so you set it once on each one and switching models brings the right level with it.

Same settings file, two models, two levels:

$ claude --model sonnet --settings '{"effortLevel":"high","modelSettings":{
    "claude-sonnet-5":{"effortLevel":"low"},
    "claude-opus-5":{"effortLevel":"xhigh"}}}' -p "/whicheffort"
EFFORT=low

$ claude --model opus   (same --settings)
EFFORT=xhigh

I never touched /effort and never edited the file between those two lines. I switched models. (/whicheffort is the six-line skill from the precedence tip: still the only way to read the real level from a script.)

Where the level is stored

Open your user settings and look at the two keys that decide this:

$ jq '{effortLevel, modelSettings}' ~/.claude/settings.json
{
  "effortLevel": "high",
  "modelSettings": {
    "claude-opus-5": { "effortLevel": "high" },
    "claude-fable-5": { "effortLevel": "high" },
    "claude-fable-5-1": { "effortLevel": "high" }
  }
}

I wrote the effortLevel at the top months ago. The three entries below it wrote themselves: /effort added one every time I confirmed a level with Enter while sitting on a different model. And since all three say what the global key already said, I have spent months on a single reasoning level across three models I use for jobs that have nothing in common. Writing this tip is what made me open the file.

Two details worth reading closely:

• The model's entry beats the global key inside the same file. That is what the first line of the test above proves: effortLevel: "high" and still EFFORT=low, because claude-sonnet-5 said low.
• It is saved under the canonical name (claude-opus-5) and still applies when you launch with the alias: that test ran as --model sonnet against a claude-sonnet-5 entry.

And one thing that does not change: another model's entry leaves you alone. With claude-opus-5 at xhigh while you run Sonnet, you get whatever Sonnet says.

How to leave it set

1. See where you are right now

/effort with no argument reports the level for the model you are on. With something saved, Current effort level: high. With nothing saved, Effort level: auto (currently high), which is the model's own default.

2. Save that model's level

/effort medium

Claude Code answers Set effort level to medium (saved as your default for new sessions) and writes the entry. The /effort slider does the same on Enter, and so does the effort slider inside /model. Want it only for this session? Press s in the slider: the answer changes to (this session only) and nothing is written. It is the same Enter / s split /model uses for the model itself.

One exception: on Fable 5, Opus 4.8 and Opus 4.7, the model holds its own default until you confirm a level by hand. Enter (or typing /effort <level>) releases that hold; launching with --effort does not. Opus 5 and Fable 5.1 have no hold.

3. Do it again on the next model

/model to switch, /effort <level> to pin it there. Which model deserves which is the Sonnet 5, Opus 5 and Fable conversation: the model that delegates and verifies takes a low level far better than the one that investigates on its own.

4. If you change your mind

/effort auto

That clears the saved level for the active model and hands it back to its default. Other models' entries and the global key stay put.

Reference

The per-model cap: maxEffortLevel

The model's entry takes one more field: maxEffortLevel. It trims anything you ask for above it, the launch flag included.

$ claude --model sonnet --effort xhigh \
    --settings '{"modelSettings":{"claude-sonnet-5":{"maxEffortLevel":"low"}}}' \
    -p "/whicheffort"
Effort 'xhigh' exceeds the cap for claude-sonnet-5 set by your settings or organization; using 'low'
EFFORT=low

That warning is the exception: effort in settings fails silently when you type a level that does not exist, and this cap is the one piece that tells you it trimmed what you asked for.

The model your subagents run on

Your subagents do not inherit your level by default. They run on the model Claude passes when it starts them, or the one in their definition, and only when neither applies do they inherit your conversation's. CLAUDE_CODE_SUBAGENT_MODEL sits behind both of those: since v2.1.251 it is a default value, not an override. To make it win outright, pair it up:

{
  "env": {
    "CLAUDE_CODE_SUBAGENT_MODEL": "haiku",
    "CLAUDE_CODE_SUBAGENT_MODEL_FORCE": "1"
  }
}

Where you check that it took is the /tasks panel. Each subagent's row carries the model it runs on, and adds the effort level when its definition, or the skill it forked from, sets effort in frontmatter. The row as the panel prints it:

craft:review › Adversarial review of cache-marker patch
9m 21s · 118.4k tokens · 45 tools · Opus 5 (1M context)

One old reason to leave the dial alone mid-task is still around: changing effort used to break the cache, and the next request reread the whole conversation. On Fable 5.1 with an API key or a Claude subscription it no longer does, as of v2.1.260: Claude Code applies the new level without asking. Every other model still asks before breaking it, which is exactly why the level is better set in advance.

Official docs: Model configuration: adjust effort level · modelSettings · Sub-agents: choose a model

Requirements

• Checked on Claude Code v2.1.269. What I ran are the --settings probes in this tip, the cap warning included. The /effort, s and /effort auto replies are the exact strings from the v2.1.269 binary rather than a captured interactive session, and the /tasks row comes from the panel itself.
• modelSettings needs v2.1.251, the slider's s key needs v2.1.257, a per-model maxEffortLevel needs v2.1.267, and the model on the /tasks row needs v2.1.242.
• Saving a level with Enter is an interactive session on your own machine: under -p, /effort applies to that session only and writes nothing.