Anthropic Just Revealed How to Prompt Fable 5.2

Summary

The video explains when to use Claude Fable 5.2 instead of Opus 5.5, compares the models on three work tasks, and presents eight prompting practices drawn from Anthropic’s guidance and the presenter’s own testing.

Performance and cost comparison

Fable 5.2 is described as up to 30% faster than its predecessor and notably stronger on some demanding tasks. However, it costs about twice as much per token as Opus 5.5: $8 per million input tokens and $40 per million output tokens, compared with Opus’s $4 and $20. The presenter also says Fable can use subscription limits quickly.

In three side-by-side tests using the same prompts and high-effort settings:

• Video editing and motion graphics: Fable produced cleaner work that better matched the supplied style, but took 51 minutes and cost $35. Opus took 35 minutes and cost $21.
• Landing-page redesign: Fable produced a more developed, visually impressive page, with more scrolling animations. It took 38 minutes and cost $18. Opus took 22 minutes and cost $10.
• Marketing presentation: Results were broadly similar. Fable’s report felt more finished, while Opus was more creative but made more errors in marketing metrics. Fable took 32 minutes and cost $34. Opus took 21 minutes and cost $20.

Overall, the three Fable runs cost $87 versus $51 for Opus—about 70% more. The presenter recommends reserving Fable for high-stakes, complex, open-ended work, such as extensive research, financial modeling, intricate visual tasks, or audits. For everyday writing and document editing, Opus is presented as the faster, cheaper default. Before switching to Fable after a disappointing result, Anthropic’s advice is to try raising the effort level first.

Eight prompting recommendations

1. Choose the model before starting a chat. Switching models mid-conversation can discard the first model’s internal analysis and require the new model to reread the conversation. The presenter also warns that model-specific caches are lost when switching.
2. Set the effort level deliberately. The video describes low effort for short exchanges, medium for routine work, high for complex research and long tasks, and extra-high or max for rare tasks where time and cost matter less. The presenter favors high effort for Fable’s intended use and notes that effort can be changed without changing models.
3. Give the model the whole task upfront. Use a four-part prompt: describe the complete task, explain its purpose and audience, specify constraints and exceptions, and define the expected deliverable and stopping point. The video stresses that Fable follows rules closely, so exceptions should be explicit.
4. Request progress updates on long tasks. Add an instruction such as: “Every few tool calls, give the user a short status report.” The presenter says Anthropic’s tests found this reduced long periods without updates.
5. Ask for concise answers. To reduce verbosity, request writing for “an intelligent but uninformed reader,” or configure a short, plain-language output style.
6. Tell it to finish the work. Anthropic’s suggested instruction asks the model to persist through common obstacles and not stop while work remains. The presenter says it reduced unfinished sessions in testing, though it used roughly 30% more tokens.
7. Audit old instructions and settings. Longstanding memory, project, or skill rules may unnecessarily constrain newer models. The presenter recommends reviewing them for outdated or overly rigid instructions.
8. Request evidence, not private reasoning. Instead of asking for the model’s internal thought process, ask for verifiable sources—for example, the exact source and lines supporting each figure in a financial report.

Main speakers and sources

• Ben, the video’s presenter, reports the model comparisons and provides the prompting advice.
• Anthropic’s Fable 5.2 prompting guide and cited testing inform several recommendations.
• Uber and Oracle are mentioned by the presenter as examples of organizations using the model for lengthy audits or extracting business indicators.