Calibration
MathVision → ZeroBench
ZeroBench is estimated from MathVision with a Hill curve fitted on 5 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (1.21665^6.00 + x^6.00), R² = 0.42, cross-validated error 4.7 pp. It is used for 14 estimates.
| Estimated model | MathVision | ZeroBench | Source |
|---|---|---|---|
| Claude Opus 4.5 | 74.3% | 5.9% | estimated ± 4.7 pp, low confidence |
| Gemini 3 Pro | 86.6% | 13.8% | estimated ± 4.7 pp, low confidence |
| Gemma 4 12B | 79.7% | 8.8% | estimated ± 4.7 pp, low confidence |
| GPT-5.2 | 83.0% | 11.0% | estimated ± 4.7 pp, low confidence |
| Kimi K2.6 | 87.4% | 14.5% | estimated ± 4.7 pp, low confidence |
| Qwen3.5-122B-A10B | 86.2% | 13.5% | estimated ± 4.7 pp, low confidence |
| Qwen3.5-27B | 86.0% | 13.3% | estimated ± 4.7 pp, low confidence |
| Qwen3.5-35B-A3B | 83.9% | 11.7% | estimated ± 4.7 pp, low confidence |
| Qwen3.5 397B | 88.6% | 15.6% | estimated ± 4.7 pp, low confidence |
| Qwen3.6 Plus | 88.0% | 15.0% | estimated ± 4.7 pp, low confidence |
| Qwen3.7 Plus | 90.3% | 17.2% | estimated ± 4.7 pp, low confidence |
| Qwen3.8-27B | 90.0% | 16.9% | estimated ± 4.7 pp, low confidence |
| Qwen3.8-Flash-Next | 90.6% | 17.5% | estimated ± 4.7 pp, low confidence |
| Qwen3.8-Omni-Flash | 91.8% | 18.7% | estimated ± 4.7 pp, low confidence |