Fable 5.2 vs Opus 5.5 and When to Use Each
Use Fable 5.2 for ambitious, open-ended work; use Opus 5.5 for well-defined tasks where you know what you want.
- Reserve Fable 5.2 for complex projects that benefit from long, autonomous runs—not routine tasks.
- Start at medium or high effort; use extra high for the hardest work, since higher effort increases cost.
- Set approval boundaries, ask it to critique your idea first, and tell it when it owns the codebase.
The notes capture the model comparison and practical prompting guidance. · Best for: Professionals choosing AI models for coding, research, and complex project work.
The episode reviews Fable 5.2’s benchmark results, costs, early use cases, and limitations. Its main practical distinction: use it for ambitious projects with room to work autonomously, while Opus 5.5 may suit specific, clearly defined tasks better.
Benchmarks, costs, and effort settings2:40
- The host reports Fable 5.2 leading several coding, science, and overall capability benchmarks, while results in agentic knowledge work were less clearly differentiated.
- The Remote Labor Index score cited is 41.3%, compared with 21% for Opus 5.5.
- The episode says Fable 5.2 costs more per task at higher effort settings. Anthropic recommends medium by default, high for everyday coding, extra high for long-running agentic work, and max for testing the model’s ceiling.
Examples of ambitious work6:24
- Anthropic’s examples include migrating a production service from Python to Rust and coordinating agents to build a prototype operating system.
- The episode also describes reported work in theoretical physics, protein structure prediction, formal proof, and robotic-arm control.
- The suggested pattern is to use Fable for difficult projects that other models cannot handle easily.
Early impressions and limitations13:30
- Early testers reported strong results in security bug finding, 3D rendering, video generation, and complex projects.
- One tester also reported errors in colors, units, and slide layout, illustrating that complex output still needs review.
- The host and other reviewers describe uneven judgment: Fable can complete difficult work but may miss what people will find useful, engaging, or good enough.
Where the model may create new opportunities20:12
- The host found that a coding harness with tools produced a more compelling capability demonstration than the regular chat interface.
- The episode argues that the model may be most valuable for creating things that were previously impractical, including functional games and custom software.
- For many ordinary uses, reviewers said Opus 5.5 may be just as suitable or less frustrating.
Choosing a model and prompting it26:51
- One suggested distinction is Opus 5.5 for specific tasks with a clear target and Fable 5.2 for creative, open-ended work.
- Give Fable a whole project to work on, specify where it needs approval, and let it make other decisions.
- For coding, say the codebase is its own; the episode says this can encourage it to consolidate repeated code and make related fixes.
Step by step
- Choose Fable 5.2 when the task is ambitious, open-ended, or difficult for other models; choose Opus 5.5 when requirements are specific.
- Ask Fable to identify weaknesses in your idea before it starts building.
- State exactly which actions require your approval, such as deleting files, and leave other decisions to the model.
- For coding, tell it that it owns the codebase.
- Start at medium or high effort; reserve extra high for the hardest, long-running work.
- Give it the whole project and review the finished work for errors in details and judgment.
Tools & resources mentioned
- Fable 5.2 — Anthropic model discussed for ambitious, open-ended tasks.
- Opus 5.5 — Anthropic model discussed as an option for specific, clearly defined tasks.
- Claude Code — Coding environment used in the host’s test of a more compelling model-generated project.
- Anthropic prompting guide — Updated guidance mentioned for prompting its models.
Notes are AI-generated from the video's transcript and may miss nuance.