benchgap
Calibration

ZeroBench → ScreenSpot Pro

ScreenSpot Pro is estimated from ZeroBench with a Hill curve fitted on 5 models measured on both: y = 0.0000 + (0.8503 − 0.0000)·x^6.00 / (0.11775^6.00 + x^6.00), R² = 0.98, cross-validated error 3.6 pp. It is used for 6 estimates.

Estimated modelZeroBenchScreenSpot ProSource
Muse Spark Open50.0%85.0%estimated ± 3.6 pp, low confidence
Kimi K3.1 Pro39.0%85.0%estimated ± 3.6 pp, medium confidence
Kimi K323.0%83.5%estimated ± 3.6 pp, medium confidence
dots3-note Preview19.0%80.5%estimated ± 3.6 pp, medium confidence
Seed 2.1 Pro18.0%78.8%estimated ± 3.6 pp, medium confidence
Seed 2.1 Turbo11.0%34.0%estimated ± 3.6 pp, low confidence

Fitted on

ModelZeroBenchScreenSpot Pro
Gemini 3.1 Pro29.0%84.4%
GPT-5.441.0%85.4%
Inkling 2 Mini17.0%76.5%
Muse Spark33.0%84.1%
Qwen3.8 Max24.0%84.5%