benchgap
Calibration

OSWorld-Verified → DeepSearchQA

DeepSearchQA is estimated from OSWorld-Verified with a Hill curve fitted on 7 models measured on both: y = 0.6926 + (1.2000 − 0.6926)·x^6.00 / (0.89987^6.00 + x^6.00), R² = 0.38, cross-validated error 7.9 pp. It is used for 10 estimates.

Estimated modelOSWorld-VerifiedDeepSearchQASource
Qwen3.8 Max86.1%91.3%estimated ± 7.9 pp, low confidence
Holo3-35B-A3B82.6%88.2%estimated ± 7.9 pp, low confidence
MiMo-V2.6-Pro82.0%87.7%estimated ± 7.9 pp, low confidence
MiMo-V2.6-Flash80.8%86.7%estimated ± 7.9 pp, low confidence
Holo3-122B-A10B78.9%85.1%estimated ± 7.9 pp, low confidence
UI-Mate-27B77.0%83.6%estimated ± 7.9 pp, low confidence
Claude Opus 4.566.3%76.3%estimated ± 7.9 pp, low confidence
UI-Mate-9B66.2%76.2%estimated ± 7.9 pp, low confidence
GPT-5.3 Codex64.7%75.4%estimated ± 7.9 pp, low confidence
Claude Sonnet 4.561.4%73.9%estimated ± 7.9 pp, low confidence

Fitted on

ModelOSWorld-VerifiedDeepSearchQA
Claude Opus 4.672.7%73.7%
Claude Opus 4.883.4%93.1%
GPT-5.475.0%73.6%
Kimi K2.673.1%92.5%
Kimi K2.563.3%77.1%
Muse Glimmer 30B65.9%74.6%
Muse Spark 1.180.8%84.9%