Fable 5.2 Is Insane... Only If You Use It Anthropic's Way

Overview

The video presents nine recommendations for using Fable 5.2 effectively in Claude Code, drawing on Anthropic guidance and the presenters’ own tests. Its central point is that Fable 5.2 is capable of demanding, long-running work, but it can consume usage limits and cost more than Opus 5.5. Use it selectively and configure the workflow around the task.

The video reports that Fable 5.2 costs $8 per million input tokens and $40 per million output tokens—20% less than Fable 5.1—and can generate responses up to 30% faster while using fewer tokens for comparable work. In Anthropic’s FrontierCode test, Fable 5.2 scored 49.1%, compared with 32.4% for Fable 5.1 and 45.5% for Opus 5.5, at the highest effort setting. The video also says Fable 5.2 costs about twice as much per token as Opus 5.5 and can exhaust usage limits faster.

The Nine Rules

1. Start with Opus 5.5, not Fable

Use Opus for everyday tasks, and move to Fable when Opus cannot handle the work, even at higher effort.

In one comparison cited in the video, Opus completed a 3D-model task in 27 minutes, while Fable took 58 minutes and used roughly seven times as many tokens for a similar result. The presenters also found that Fable cost about 58% more on a feature they tested.

2. Define what “done” means before starting

Give the model measurable completion criteria rather than vague requests such as “make the page pretty.” Break the work into smaller tasks, each with a testable pass/fail condition, and set a high quality bar.

For review, the presenters recommend having Fable assess the finished work in a separate session. They found that it identified issues the builder model had missed.

3. Reserve Fable for especially suitable work

The video recommends Fable for:

• Large-scale work, such as restructuring an application or rewriting substantial parts of a program.
• 3D graphics and game development, including building scenes and iterating on complex results.
• The most difficult coding problems, where its stronger performance may justify the extra cost.

4. Use Fable as a lead agent and delegate selectively

Fable can plan work, divide it into subtasks, assign those to other models, and review their results. The suggested division is:

• Sonnet for straightforward tasks.
• Opus 5.5 for more complex work.
• Haiku for small jobs, such as searching files or reading logs.

Because subagents add cost, limit them to work that can genuinely run in parallel or to large projects. For a small, single-session feature, delegation may cost more than it saves. The video suggests documenting these model-selection rules in Claude.md.

5. Use high effort for long-running lead-agent work

Anthropic reportedly found that Fable may stop before completing a long task, particularly at low or medium effort. For unattended work, set the lead agent to high effort.

Alternatively, at medium effort, add Anthropic’s continuation instructions to Claude.md. The video says this reduced incomplete sessions in Anthropic’s tests, though it did not address every reason the model might stop. Avoid going above high effort unless cost and time are unimportant.

6. Set up permissions before leaving a task unattended

Configure Claude Code’s auto mode and document which project folders, repositories, and test environments Fable may access.

The video’s rule of thumb is that the model can handle reversible actions independently, while actions that cannot be undone should wait for approval. When approving a request, specify exactly what is allowed rather than replying only “yes.”

7. Require a useful final report

Write instructions so the report makes sense to someone who did not observe the work: put the main result first and use plain language. Ask for evidence, such as screenshots and test results, since a text summary may omit small changes. The video also suggests asking for a browser-viewable report when helpful.

8. Audit outdated Claude Code instructions

Old Claude.md rules or skills may confuse a newer model. In particular, wording such as “never” can be interpreted as an absolute restriction.

The video suggests phrasing approval requirements more precisely—for example, “ask before deleting files”—and using Claude Code’s prompt audit command to check instructions and skills against the model in use.

9. Plan model switches to avoid losing context

Claude reportedly keeps a model-specific copy of a conversation for an hour, so switching models can require the new model to reread the conversation, increasing token use. The video also warns that Fable’s reasoning before its answers does not transfer to the new model.

When the current task is complete, compress the conversation or begin a new session before switching.

Practical Checklist

• Choose the model based on task difficulty: begin with Opus 5.5 and escalate when needed.
• Write measurable completion criteria and task-level tests.
• Set Fable as a lead agent in Claude Code and assign simpler subtasks to less costly models.
• Limit subagents to substantial or parallelizable work.
• Set high effort for unattended long tasks, or use continuation instructions at medium effort.
• Configure auto mode, allowed locations, and approval rules in advance.
• Request evidence-based, plain-language completion reports.
• Run prompt audit to review older project instructions and skills.
• Compress context or start a new session before changing models.

Speakers and Sources

• AI Labs presenter/narrator, reporting the channel’s tests and recommendations.
• Lydia Hallie, whose Claude Code best-practices talk is cited.
• Addy Osmani, whose guide on choosing Claude models is cited.
• Anthropic documentation and testing, including guidance on Fable, effort settings, and model behavior.
• Unity and an X user/developer, cited for examples involving 3D and game development.