benchgap
Calibration

ZeroBench → PerceptionBench

PerceptionBench is estimated from ZeroBench with a offset logistic curve fitted on 5 models measured on both: y = 0.5545 + (0.6830 − 0.5545) / (1 + exp(−169.73·(x − 0.2369))), R² = 0.94, cross-validated error 2.7 pp. It is used for 6 estimates.

Estimated modelZeroBenchPerceptionBenchSource
Muse Spark Open50.0%68.3%estimated ± 2.7 pp, low confidence
GPT-5.441.0%68.3%estimated ± 2.7 pp, low confidence
Muse Spark33.0%68.3%estimated ± 2.7 pp, medium confidence
Gemini 3.1 Pro29.0%68.3%estimated ± 2.7 pp, medium confidence
Seed 2.1 Pro18.0%55.5%estimated ± 2.7 pp, medium confidence
Seed 2.1 Turbo11.0%55.5%estimated ± 2.7 pp, low confidence

Fitted on

ModelZeroBenchPerceptionBench
dots3-note Preview19.0%53.4%
Inkling 2 Mini17.0%57.5%
Kimi K3.1 Pro39.0%68.3%
Kimi K323.0%58.5%
Qwen3.8 Max24.0%63.5%