跪拜 Guibai
← All articles
Claude

Fable 5.2: Make It the Brain, Not the Wallet-Burner

By XiaoLei_Liu ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

On a $20 Claude Pro plan, running Fable on routine chores can drain the weekly quota in a single afternoon, and after November 16 Pro plans consume usage credits for Fable. The concrete payoff here is a config that offloads grunt work to cheap models and only spends Fable's $40-per-million output tokens where it actually beats Opus.

Summary

Fable 5.2's cost structure punishes misrouting: $8/$40 per million input/output tokens makes it 80x more expensive than Haiku per token, yet Anthropic's own tests show it beats Opus on the hardest benchmarks at roughly half the cost per task. The practical side of the model, laid out with real configuration prompts, is knowing when to spend those tokens. The guide's core move is a model ladder: Fable plans and does root-cause debugging, Opus builds, Sonnet reviews, Haiku handles renaming and other easy-to-verify chores.

Around that ladder sit two cost-critical Claude Code features. Advisor mode lets Sonnet run the session and consult Fable only at decision points, and it preserves the prompt cache, unlike switching /model mid-session. For larger jobs, a four-agent setup with `/brain` on Fable, plus subagents on Opus, Sonnet, and Haiku, is claimed to cut the cost of a hard bug hunt from $563 to $165 while barely dropping the score. The guide also includes a prompt doctor that scans existing CLAUDE.md files, skills, and agents for instructions written for older Claude models that now waste money or reduce quality.

The biggest operational wins are concrete: a "finishing work" instruction halves premature stops at a 30% token surcharge; writing exceptions directly into approval rules lifts execution from 1-in-10 to 9-in-10; and the advisor no-op trap is real, 0 consultations in 198 problems when the prompt doesn't say when to ask. Beyond cost, there are batteries-included routines for a personal mirror analysis, a Karpathy-style wiki second brain, a bill autopsy, and an unattended overnight task run with a `/goal` guardrail.

Takeaways
— Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10/$50, with a fast mode at $16/$80.
— Fable and Mythos 5.2 are the same underlying model; Mythos removes safety guardrails and is restricted to approved organizations.
— Anthropic's model ladder assigns Fable to planning, architecture, and root-cause debugging; Opus to complex reasoning; Sonnet to daily coding; Haiku to simple high-volume chores.
— On FrontierCode 2.0, Fable at high effort scores 45.3% versus Opus 5.5's 45.5% at absolute limit, but at roughly half the cost per task; at xhigh Fable hits 49.1%.
— The 'finishing work' CLAUDE.md module halves the number of sessions that stop early, at a cost of about 30% more output tokens.
— Writing an exception into an approval rule (e.g. 'get approval before X') makes Fable execute 9 out of 10 times after approval; a 'never' rule plus a separate approval note only gets 1 out of 10.
— The Advisor mode trap: in Anthropic's test, the advisor was consulted 0 out of 198 problems unless the prompt explicitly said when to consult it.
— A Fable-as-brain plus Sonnet-as-hands configuration found a planted bug in a 22-million-token codebase for $165, 71% cheaper than Fable alone at $563, with slightly lower score; Opus as brain was $139 with similar score.
— Toggling /advisor preserves the prompt cache, while switching /model midway invalidates it; Fable reads cached input at 1/40 the normal input price.
— Running low effort first and rerunning only failures on high saves about 40%, and providing data as files instead of pasting into chat raised accuracy from 7/25 to 25/25 at 1/12 the cost.
— Claude Code's /goal command uses a separate model to verify completion after each turn, and --permission-mode auto has a second model review operation safety.
— Fable's security guardrails block most cyber and bio-sensitive work, and Anthropic deliberately kept its ability to speed up AI research itself at Opus 5.5 level.
— After November 16, the $20 Pro plan requires usage credits for Fable, while Max and Ultra include up to half the weekly limit.
Conclusions

The 80x per-token gap between Fable and Haiku makes model routing a bigger financial lever than prompt phrasing; the article's whole cost strategy turns on that one ratio.

Anthropic's own tests show that a Fable brain plus cheap hands is not a compromise, it's often the rational default, because the highest-cost model contributes the thinking, not the typing.

The strongest behavioral effect comes from a plain-English instruction to finish work before stopping, which suggests that Fable's tendency to stop early is more of a product default than a reasoning failure.

The advisor is functionally invisible unless the prompt gives it a schedule; a 0-out-of-198 consultation rate shows that adding a model feature is not the same as making it useful.

Concepts & terms
Model ladder
Anthropic's recommended routing of Claude Code tasks across four models: Fable for planning and hard debugging, Opus for complex reasoning, Sonnet for daily coding, Haiku for simple high-volume chores.
Advisor mode
A Claude Code setting where a cheaper model runs the session and calls on Fable for guidance at key decision points, saving cost while keeping Fable's judgment in the loop.
Effort levels
Claude Code settings (medium, high, xhigh) that control how much reasoning Fable spends on a task. Most capability gain shows up at high; xhigh is for the hardest long-running problems.
ultracode
A Claude Code built-in command that automatically writes a workflow to spawn many parallel agents in the background and aggregate their results into a report.
/deep-research
A Claude Code command that performs broader web search, cross-verifies sources, and drops weak claims before producing a report.
CLAUDE.md
A memory file that holds persistent instructions and preferences for Claude Code sessions, remembered across projects at user or project level.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗