Claude Fable 5.5 (Leaked Checkpoint) vs Gemini 4 Pro (Leaked Benchmark)
A direct, empirical evaluation of Claude Fable 5.5 (Leaked Checkpoint) (Anthropic) against Gemini 4 Pro (Leaked Benchmark) (Google DeepMind).
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.
Gemini 4 Pro (Leaked Benchmark)
Developed by Google DeepMind · Proprietary Omni-Modal Unified Diffusion Transformer parameters
Repository-scale 2M-token refactoring (88.7% DeepSWE v1.1), headless shell execution (95.3% Terminal-Bench 2.1), and multimodal desktop agent autonomy (86.8% OSWorld 2.0).
| Metric | Claude Fable 5.5 (Leaked Checkpoint) | Gemini 4 Pro (Leaked Benchmark) |
|---|---|---|
| SWE-bench Verified | 84.2% | 88.7% |
| MMLU-Pro | 94.6% | 93.8% |
| MATH-500 | 98.4% | 97.8% |
| Context Window | 1,500,000 tokens | 2,000,000 tokens |
| Input Pricing (per 1M) | $3.50 | $2.25 |
| Output Pricing (per 1M) | $17.50 | $11.25 |