Fable 5.2 effort in Claude Code: use high only for these tasks

TL;DR Keep Fable 5.2 on medium, the level Claude Code starts it at, for everyday work. For the long task you're going to leave running on its own, open /effort, pick high and press s: it applies to that session only. Use xhigh only where you've measured a gain; here it cost seven times as much for the same result.

Claude Code starts Fable 5.2 at medium, while the API starts it at high. And effort doesn't carry over from the previous model: Anthropic's Fable 5.2 guide warns that how much each level thinks can differ from one model to the next, and says to rerun your sweep instead of keeping a setting tuned on Fable 5.1 or Opus 5.5. The level you saved for Fable 5.1 doesn't reach Fable 5.2 either, as you'll see below.

Which level for which work

What Anthropic's guide recommends, level by level, with Claude Code's default marked (on the API it's high):

According to Anthropic, high already gives you most of the capability gain: the step up from medium raised scores on most of the evaluations they track. Above high, gains are uneven and depend on the task.

What each level costs on a real task

The three tasks from the Opus 5.5 effort benchmark, unchanged: closing an open redirect, fixing monthly renewals and reading CSV exports from Excel and Google Sheets correctly. No prompt lists the edge cases, and every run is scored against its task's hidden suite (45, 20 and 23 cases, 88 in all), which the model never sees. Claude Code v2.1.322, four levels, two runs per task:

claude -p "<task prompt>" --model fable --effort high \
  --setting-sources project --strict-mcp-config \
  --permission-mode acceptEdits \
  --allowedTools "Bash(php *),Read,Edit,Write,Grep,Glob" \
  --output-format json

Permissions only allow editing files and running php: when it tried another shell command it was denied and took another route, 10 times at low and 29 at xhigh, across each level's 6 runs.

Medians per run, plus the total of each level's 6 runs:

All 24 runs passed all 88 cases. On these tasks the level only changed what you pay and how long you wait: high cost almost twice as much as medium and took twice as long; xhigh cost 7.4 times as much and took almost 9 times as long. One xhigh run on the open redirect took 20 minutes of API time, 32 turns and $6.99 for the same 45 out of 45 that medium got in under two minutes.

low landed level with medium: $2.79 against $3.01 in total, a little cheaper on the open redirect and even on the other two. The guide doesn't compare those two levels, but it does say low scores close to what Fable 5.1 did at higher levels. On Opus 5.5, the same tasks at xhigh cost three times what medium did (an older Claude Code version, three repetitions and no permission limits, so compare the ratio only). On Fable 5.2 that jump is more than twice as steep.

A greeting doesn't think at any level

The guide says that at high and above the model thinks before almost every reply, even to a greeting. On Claude Code v2.1.322 that doesn't happen: "hola" at low, medium, high, xhigh and max returned the same 26 output tokens with 0 thinking, and the session log holds no thinking block. Same in an interactive session at high.

A small fix (make slugify() strip accents and repeated dashes, with a test) came out the same at all four levels: 10 out of 10 hidden cases, $0.64 to $0.71 a run, because almost all of that is writing the context to the cache. Thinking was 0 tokens at low and medium, 47 to 108 at high and about 380 at xhigh.

Lowering effort doesn't make trivial turns faster in Claude Code. The savings showed up on the benchmark tasks, one to several minutes long; on the small fix, no level saved anything.

Your Fable 5.1 level doesn't carry over

Fable 5.1 started at high in Claude Code, and /effort saves the level per model, under its canonical name. With "claude-fable-5-1": {"effortLevel": "high"} under modelSettings and a top-level effortLevel of xhigh, both in your ~/.claude/settings.json:

$ claude -p "/whicheffort" --model claude-fable-5-1
EFFORT=high

$ claude -p "/whicheffort" --model fable
EFFORT=medium

/whicheffort is the probe from the effort chain. Fable 5.2 ignores both lines: the entry belongs to another model, and a top-level effortLevel in your user settings doesn't count for models released since Opus 5.5. If claude-fable-5-1 was your saved model, on Pro, Max or Ultra it moved to the fable alias by itself (on Team and Enterprise it can come later, or not at all), and since then you've been working at medium without picking it, unless another source sets the level, such as CLAUDE_CODE_EFFORT_LEVEL or a project effortLevel.

Set it once

1. Check the level you're on

$ claude -p "/effort status" --model fable
Effort level: auto (currently medium)

auto (currently medium) means you have nothing saved for Fable 5.2 and Claude Code sends its default, medium.

2. If you want a different default, save it for Fable 5.2

With /model fable active, open /effort, move the slider with the arrow keys and confirm with Enter:

Set effort level to high (saved as your default for new sessions): Comprehensive implementation with extensive testing and documentation

It lands in your ~/.claude/settings.json, for this model only:

{
  "modelSettings": {
    "claude-fable-5-2": {
      "effortLevel": "high"
    }
  }
}

With the numbers above, the sensible move is to save nothing and stay on medium. To keep one level per model across every model you use, set effort per model.

3. For the long part, raise only that session

Plan and iterate at medium. When you're about to leave a long task running on its own, open /effort, pick high and press s: it applies to this session and leaves your saved level alone (v2.1.257 or later). If it still ends the turn before the work is done, here's what else works.

4. xhigh and max only with a measurement in hand

xhigh only if a task of yours really improves with it. max can't be saved: in Claude Code it lasts one session, unless you pin it with CLAUDE_CODE_EFFORT_LEVEL.

Official docs: Model configuration: Adjust effort level · Prompting Claude Fable 5.2: Calibrate effort · Effort

What this test doesn't measure

• Tasks that run for hours. The longest run took 20 minutes of API time. The guide puts long unattended runs at high, and that's where medium stops before finishing more often.
• Sample size. Two runs per task and level, one day, API list price. On a subscription you feel it in your limit and in the wait, and in dollars if your plan bills Fable to usage credits.

Requirements: Claude Code v2.1.321 or later for fable to be Fable 5.2 (in Claude apps gateway sessions, /model claude-fable-5-2). Checked on v2.1.322 on November 12, 2026.