Calibration
ScreenSpot Pro → V*
V* is estimated from ScreenSpot Pro with a Hill curve fitted on 5 models measured on both: y = 0.0000 + (0.9713 − 0.0000)·x^6.00 / (0.39723^6.00 + x^6.00), R² = 0.88, cross-validated error 6.5 pp. It is used for 14 estimates.
| Estimated model | ScreenSpot Pro | V* | Source |
|---|---|---|---|
| GPT-6 Astra | 92.7% | 96.5% | estimated ± 6.5 pp, low confidence |
| Claude Opus 4.8 | 87.9% | 96.3% | estimated ± 6.5 pp, low confidence |
| GPT-5.4 | 85.4% | 96.2% | estimated ± 6.5 pp, low confidence |
| Qwen3.8 Max | 84.5% | 96.1% | estimated ± 6.5 pp, low confidence |
| Gemini 3.1 Pro | 84.4% | 96.1% | estimated ± 6.5 pp, low confidence |
| Muse Spark | 84.1% | 96.1% | estimated ± 6.5 pp, low confidence |
| Claude Opus 4.6 | 83.1% | 96.0% | estimated ± 6.5 pp, low confidence |
| Qwen3.7 Plus | 79.0% | 95.6% | estimated ± 6.5 pp, low confidence |
| Muse Glimmer 30B | 75.4% | 95.1% | estimated ± 6.5 pp, low confidence |
| Holo2-235B-A22B | 70.6% | 94.1% | estimated ± 6.5 pp, low confidence |
| Holo2-30B-A3B | 66.1% | 92.8% | estimated ± 6.5 pp, low confidence |
| Holo2-8B | 58.9% | 88.8% | estimated ± 6.5 pp, low confidence |
| Nemotron 3 Nano Omni 30B A3B | 57.8% | 87.9% | estimated ± 6.5 pp, low confidence |
| Holo2-4B | 57.2% | 87.3% | estimated ± 6.5 pp, low confidence |
Fitted on
| Model | ScreenSpot Pro | V* |
|---|---|---|
| Claude Opus 4.5 | 45.7% | 67.0% |
| Gemini 3 Pro | 72.7% | 88.0% |
| LFM3-VL-2B | 55.5% | 86.4% |
| Qwen3.5 397B | 65.6% | 95.8% |
| Qwen3.6 Plus | 68.2% | 96.9% |