Claude Fable 5.5 (Leaked Checkpoint) vs DeepSeek-R1
A direct, empirical evaluation of Claude Fable 5.5 (Leaked Checkpoint) (Anthropic) against DeepSeek-R1 (DeepSeek-AI).
Claude Fable 5.5 (Leaked Checkpoint)
Developed by Anthropic · Proprietary Frontier Dynamic Reasoner parameters
Terminal-Bench 4.0 autonomous shell execution (78.5% target), recursive self-critique, and 1.5M-token repository refactoring.
DeepSeek-R1
Developed by DeepSeek-AI · 671B parameters
Olympiad mathematics, logic verification, and autonomous thinking chains.
| Metric | Claude Fable 5.5 (Leaked Checkpoint) | DeepSeek-R1 |
|---|---|---|
| SWE-bench Verified | 84.2% | 49.2% |
| MMLU-Pro | 94.6% | 84% |
| MATH-500 | 98.4% | 97.3% |
| Context Window | 1,500,000 tokens | 128,000 tokens |
| Input Pricing (per 1M) | $3.50 | $0.55 |
| Output Pricing (per 1M) | $17.50 | $2.19 |