Calibration
ScreenSpot Pro → RealWorldQA
RealWorldQA is estimated from ScreenSpot Pro with a linear curve fitted on 6 models measured on both: y = 0.2850·x + 0.6393, R² = 0.96, cross-validated error 1.0 pp. It is used for 13 estimates.
| Estimated model | ScreenSpot Pro | RealWorldQA | Source |
|---|---|---|---|
| GPT-6 Astra | 92.7% | 90.3% | estimated ± 1.0 pp, low confidence |
| Claude Opus 4.8 | 87.9% | 89.0% | estimated ± 1.0 pp, low confidence |
| GPT-5.4 | 85.4% | 88.3% | estimated ± 1.0 pp, low confidence |
| Gemini 3.1 Pro | 84.4% | 88.0% | estimated ± 1.0 pp, medium confidence |
| Muse Spark | 84.1% | 87.9% | estimated ± 1.0 pp, medium confidence |
| Claude Opus 4.6 | 83.1% | 87.6% | estimated ± 1.0 pp, medium confidence |
| Muse Glimmer 30B | 75.4% | 85.4% | estimated ± 1.0 pp, medium confidence |
| Holo2-235B-A22B | 70.6% | 84.0% | estimated ± 1.0 pp, medium confidence |
| Qwen3.6 Plus | 68.2% | 83.4% | estimated ± 1.0 pp, medium confidence |
| Holo2-30B-A3B | 66.1% | 82.8% | estimated ± 1.0 pp, medium confidence |
| Holo2-8B | 58.9% | 80.7% | estimated ± 1.0 pp, medium confidence |
| Nemotron 3 Nano Omni 30B A3B | 57.8% | 80.4% | estimated ± 1.0 pp, medium confidence |
| Holo2-4B | 57.2% | 80.2% | estimated ± 1.0 pp, medium confidence |
Fitted on
| Model | ScreenSpot Pro | RealWorldQA |
|---|---|---|
| Claude Opus 4.5 | 45.7% | 77.0% |
| Gemini 3 Pro | 72.7% | 83.3% |
| LFM3-VL-2B | 55.5% | 79.3% |
| Qwen3.5 397B | 65.6% | 83.9% |
| Qwen3.7 Plus | 79.0% | 86.9% |
| Qwen3.8 Max | 84.5% | 88.0% |