Surface Evolver Bench Pass Rate
Agents · 2026-06-04
Strict task pass rate on Surface Evolver Bench, where a task passes only when static checks and all hidden dynamic checks pass. Models write complete Surface Evolver .fe simulation datafiles using the benchmark's documentation, execution-feedback, and structured-submission tool environment.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 87.5 |
| Kimi K3 | 87.5 |
| GPT-5.6 Sol | 87.5 |
| GPT-5.5 | 75.0 |
| Opus 4.8 | 68.8 |
| GPT-5.6 Terra | 68.8 |
| Gemini 3.8 Flash | 56.3 |
| Grok 4.5 | 50.0 |
| Gemini 3.5 Flash | 37.5 |
| Muse Spark 1.1 | 37.5 |
| Kimi K2.7 Code | 37.5 |
| GLM-5.2 | 31.3 |
| DeepSeek V4.1 Flash | 31.3 |
| DeepSeek-V4-Pro | 25.0 |
| GPT-5.6 Luna | 18.8 |
| Sonnet 5 | 18.8 |
| MiniMax M3 | 18.8 |
| GLM-5.3 Flash | 18.8 |
| Qwen3.8 27B | 18.8 |
| Mistral Large 4 Preview | 18.8 |
| Qwen3.6 35B A3B | 12.5 |
| Gemma 4 31B IT | 6.3 |
| Mistral Medium 3.5 | 6.3 |
| gpt-oss-120b | 6.3 |
| Laguna M.1 | 6.3 |
| Trinity Large Thinking | 6.3 |