Claude Fable 5 review: AI model crushes benchmarks but struggles with practical GTM tasks

The Gist

• Claude Fable 5 hits 80% on SWBench Pro, outperforming GPT-4.5 and Gemini 3.1 Pro
• Costs $10/$50 per million input/output tokens, twice the rate of cheaper models
• Excels at vision tasks but produces unreadable specs and PRDs

• Claude Fable 5 excels in benchmarks (80% on SWBench Pro) but struggles with practical tasks like design and writing readable specs.
• The model is expensive ($10/$50 per million tokens) and consumes tokens at twice the rate of other models, requiring strategic deployment.
• Fable 5 is exceptionally good at vision tasks (e.g., document formatting, PDF parsing) but produces poor design output for one-shot tasks.
• The model is overly conservative in execution, often delivering minimal but not useful MVPs due to built-in safety guardrails.
• Multi-agent orchestration is technically possible but unreliable, with frequent stalls and errors.
• Match model intelligence to task complexity: use Fable 5 for hard technical problems and vision tasks, but cheaper models for front-end work and design.

• Use Fable 5 strategically for high-complexity tasks (e.g., vision, technical problems) but opt for cheaper models for simpler tasks.
• Avoid relying on Fable 5 for design or spec-writing tasks due to its poor output quality in these areas.
• Monitor multi-agent workflows closely, as they are prone to stalls and errors.
• Leverage Fable 5’s vision capabilities for document parsing and formatting tasks.

• 80% (Fable 5's score on SWBench Pro, outperforming Opus 4.8, GPT-4.5, and Gemini 3.1 Pro.)
• $10/$50 per million tokens (Cost of input/output tokens for Fable 5, making it more expensive than other models.)
• 95% (Sessions that don’t trigger a fallback to Opus 4.8 due to safety guardrails.)
• 30-day retention policy (Anthropic's policy to catch misuse of Fable 5.)