Calibration
ScreenSpot Pro → CharXiv
CharXiv is estimated from ScreenSpot Pro with a Michaelis–Menten curve fitted on 12 models measured on both: y = 1.2249·x / (0.35656 + x), R² = 0.71, cross-validated error 3.9 pp. It is used for 6 estimates.
| Estimated model | ScreenSpot Pro | CharXiv | Source |
|---|---|---|---|
| GPT-6 Astra | 92.7% | 88.5% | estimated ± 3.9 pp, medium confidence |
| Claude Opus 4.6 | 83.1% | 85.7% | estimated ± 3.9 pp, high confidence |
| Holo2-235B-A22B | 70.6% | 81.4% | estimated ± 3.9 pp, high confidence |
| Holo2-30B-A3B | 66.1% | 79.6% | estimated ± 3.9 pp, high confidence |
| Holo2-8B | 58.9% | 76.3% | estimated ± 3.9 pp, high confidence |
| Holo2-4B | 57.2% | 75.5% | estimated ± 3.9 pp, high confidence |
Fitted on
| Model | ScreenSpot Pro | CharXiv |
|---|---|---|
| Claude Opus 4.5 | 45.7% | 68.5% |
| Claude Opus 4.8 | 87.9% | 89.9% |
| Gemini 3.1 Pro | 84.4% | 80.2% |
| Gemini 3 Pro | 72.7% | 81.4% |
| GPT-5.4 | 85.4% | 82.8% |
| Muse Glimmer 30B | 75.4% | 78.8% |
| Muse Spark | 84.1% | 86.4% |
| Nemotron 3 Nano Omni 30B A3B | 57.8% | 76.3% |
| Qwen3.5 397B | 65.6% | 80.8% |
| Qwen3.6 Plus | 68.2% | 81.5% |
| Qwen3.7 Plus | 79.0% | 85.9% |
| Qwen3.8 Max | 84.5% | 93.5% |