Calibration
VulcanBench v3 → CursorBench 3.2
CursorBench 3.2 is estimated from VulcanBench v3 with a Michaelis–Menten curve fitted on 8 models measured on both: y = 1.8682·x / (1.55108 + x), R² = 0.43, cross-validated error 3.3 pp. It is used for 6 estimates.
| Estimated model | VulcanBench v3 | CursorBench 3.2 | Source |
|---|---|---|---|
| DeepSeek V4 Flash 0731 | 88.4% | 67.8% | estimated ± 3.3 pp, medium confidence |
| Muse Spark 1.2 | 87.0% | 67.1% | estimated ± 3.3 pp, medium confidence |
| Qwen3.8-27B | 82.6% | 64.9% | estimated ± 3.3 pp, medium confidence |
| Qwen3.8 Max | 81.2% | 64.2% | estimated ± 3.3 pp, medium confidence |
| GLM-5.3 | 78.3% | 62.7% | estimated ± 3.3 pp, medium confidence |
| Claude Haiku 4.5 | 76.2% | 61.5% | estimated ± 3.3 pp, medium confidence |
Fitted on
| Model | VulcanBench v3 | CursorBench 3.2 |
|---|---|---|
| Claude Fable 5 | 89.5% | 70.5% |
| Claude Opus 5 | 87.0% | 70.0% |
| GPT-5.6 Luna | 85.5% | 61.1% |
| GPT-5.6 Sol | 87.0% | 67.2% |
| GPT-5.6 Terra | 87.0% | 64.9% |
| Grok 4.5 | 89.9% | 66.7% |
| Grok 4.6 | 87.0% | 70.8% |
| Kimi K3 | 73.7% | 60.8% |