Anthropic released the Claude Fable 5 model on June 9. After two days of developer testing, its 80.3% SWE-Bench Pro benchmark has been verified in actual coding tasks. Developers report significant improvements over Opus 4.8 in complex multi-file refactorings, autonomous debugging sessions, and large codebase analysis. The model's "long autonomous work" ability (which can continue to work for longer periods of time without human input) has become the biggest highlight.
Core technical features of Fable 5
| Characteristics | Description | Developer feedback |
| SWE-Bench Pro | 80.3%(highest in the industry) | The accuracy of multi-file reconstruction has been significantly improved |
| Over-long autonomous work | Longer continuous work than any predecessor Claude model | Reduced frequency of long session interruptions |
| secure routing | Cybersecurity/biology queries automatically transfer to Opus 4.8 | The trigger rate was less than 5%, in line with expectations |
| free period | Pro/Max/Team/Enterprise plan is free until June 22 | Developers actively try |
| pricing | Points are required from June 23 | Specific prices have not yet been announced |
Key considerations for API developers
There are important changes in Fable 5's API behavior: When the model refuses to answer for security reasons, it returns an HTTP 200 status code (instead of an error code) with a field called 'stop_reason: "refuse". If no output is generated, no charge will be paid. Developers need to handle this "soft rejection" scenario gracefully.
"Programming Three Kingdoms Kill" with GPT-5.5 and Gemini 3.5 Pro
The release of Fable 5 brings the AI programming model competition into a new stage:
| model | Company | SWE-Bench Pro | Core advantage |
| Claude Fable 5 | Anthropic | 80.30% | Long-term independent work, corporate safety and compliance |
| GPT-5.5 | OpenAI | ~58.6% | The largest developer ecosystem, Codex integration |
| Gemini 3.5 Pro | undisclosed | Deep integration with Google Cloud for cost advantages |
Claude Fable 5's 80.3% SWE-Bench Pro verification is a "new ceiling" for AI programming capabilities. While GPT-5.5 is still hovering around 58%, Anthropic has pushed its benchmark to 80%-which means AI can solve more than 4/5 of real software engineering problems autonomously. The ability to "long autonomous work" is particularly critical: it allows AI to transition from "assistive tools"(requiring human confirmation each time) to "autonomous agents"(working independently for a long time). But feedback from the developer community also reveals risks: Fable 5's "soft rejection" mechanism protects security, but also increases the processing complexity of developers; and the pricing after the free period ends on June 22 will determine whether it can transform from "technological leadership" to "commercial success." For Anthropic, Fable 5 is not only a model release, but also a key bargaining chip to showcase the "technology moat" to investors on the eve of its IPO.