benchgap
Calibration

SkillsBench official leaderboard (BenchFlow) → EBR-bench

EBR-bench is estimated from SkillsBench official leaderboard (BenchFlow) with a linear curve fitted on 10 models measured on both: y = 0.8477·x + -0.2958, R² = 0.45, cross-validated error 9.1 pp. It is used for 10 estimates.

Estimated modelSkillsBench official leaderboard (BenchFlow)EBR-benchSource
GLM-5.158.4%19.9%estimated ± 9.1 pp, low confidence
Hy355.9%17.8%estimated ± 9.1 pp, low confidence
Gemini 3 Flash54.6%16.7%estimated ± 9.1 pp, low confidence
MiniMax M353.0%15.4%estimated ± 9.1 pp, low confidence
Claude Sonnet 4.647.2%10.4%estimated ± 9.1 pp, low confidence
Grok 4.341.7%5.8%estimated ± 9.1 pp, low confidence
GPT-5.4 mini41.4%5.5%estimated ± 9.1 pp, low confidence
MiniMax M2.734.9%0.0%estimated ± 9.1 pp, low confidence
Claude Haiku 4.530.1%0.0%estimated ± 9.1 pp, low confidence
Gemini 3.1 Flash-Lite20.1%0.0%estimated ± 9.1 pp, low confidence

Fitted on

ModelSkillsBench official leaderboard (BenchFlow)EBR-bench
Claude Opus 4.549.0%14.3%
Claude Opus 4.650.2%12.7%
Claude Opus 4.761.2%19.0%
Claude Opus 4.854.1%28.6%
Claude Sonnet 4.536.2%2.4%
Gemini 3.1 Pro60.8%14.3%
Gemini 3.5 Flash48.2%4.8%
GPT-5.251.7%23.0%
GPT-5.567.3%34.3%
Kimi K2.654.0%2.4%