Calibration
ScreenSpot Pro → ParseBench
ParseBench is estimated from ScreenSpot Pro with a Michaelis–Menten + offset curve fitted on 6 models measured on both: y = 0.0127 + 2.0000·x / (1.91088 + x), R² = 0.66, cross-validated error 5.5 pp. It is used for 4 estimates.
| Estimated model | ScreenSpot Pro | ParseBench | Source |
|---|---|---|---|
| Holo2-235B-A22B | 70.6% | 55.2% | estimated ± 5.5 pp, low confidence |
| Holo2-30B-A3B | 66.1% | 52.7% | estimated ± 5.5 pp, low confidence |
| Holo2-8B | 58.9% | 48.4% | estimated ± 5.5 pp, low confidence |
| Holo2-4B | 57.2% | 47.3% | estimated ± 5.5 pp, low confidence |
Fitted on
| Model | ScreenSpot Pro | ParseBench |
|---|---|---|
| Claude Opus 4.6 | 83.1% | 54.1% |
| Claude Opus 4.8 | 87.9% | 62.3% |
| Gemini 3.1 Pro | 84.4% | 69.1% |
| GPT-5.4 | 85.4% | 62.2% |
| GPT-6 Astra | 92.7% | 69.9% |
| Nemotron 3 Nano Omni 30B A3B | 57.8% | 48.5% |