Claude Sonnet 5.5 vs Gemini 4 Pro (Leaked Benchmark)
A direct, empirical evaluation of Claude Sonnet 5.5 (Anthropic) against Gemini 4 Pro (Leaked Benchmark) (Google DeepMind).
Claude Sonnet 5.5
Developed by Anthropic · Proprietary Frontier Dynamic Hybrid Reasoner parameters
Terminal-Bench 4.0 execution (70.6%), SWE-bench Verified coding (82.4%), GitHub Copilot multi-file edits, and low-latency autonomous agentic coding at $2/$10 pricing.
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 Sonnet 5.5 | Gemini 4 Pro (Leaked Benchmark) |
|---|---|---|
| SWE-bench Verified | 82.4% | 88.7% |
| MMLU-Pro | 88.4% | 93.8% |
| MATH-500 | 95.2% | 97.8% |
| Context Window | 1,000,000 tokens | 2,000,000 tokens |
| Input Pricing (per 1M) | $2.00 | $2.25 |
| Output Pricing (per 1M) | $10.00 | $11.25 |