benchgap
Calibration

ResearchClawBench → GDPval-AA

GDPval-AA is estimated from ResearchClawBench with a Michaelis–Menten curve fitted on 13 models measured on both: y = 2.0000·x / (0.87316 + x), R² = 0.39, cross-validated error 8.3 pp. It is used for 3 estimates.

Estimated modelResearchClawBenchGDPval-AASource
MiMo-V2.516.9%32.4%estimated ± 8.3 pp, low confidence
MiMo-V2-Pro15.3%29.8%estimated ± 8.3 pp, low confidence
Grok 4.113.5%26.8%estimated ± 8.3 pp, low confidence

Fitted on

ModelResearchClawBenchGDPval-AA
Claude Opus 4.821.1%47.7%
Gemini 3.1 Pro13.3%14.6%
Gemini 3.5 Flash18.0%42.2%
GLM-5.118.2%30.8%
GLM-5.220.7%43.6%
GPT-5.415.3%37.3%
GPT-5.517.0%42.5%
Grok 4.312.4%29.2%
Kimi K2.618.0%26.9%
Kimi K2.514.0%17.0%
MiniMax M319.8%37.1%
Qwen3.6 Plus18.0%24.5%
Qwen3.7 Max18.7%31.4%