benchgap
Calibration

WideResearch → JobBench

JobBench is estimated from WideResearch with a linear curve fitted on 5 models measured on both: y = 4.4991·x + -3.1576, R² = 0.98, cross-validated error 5.1 pp. It is used for 8 estimates.

Estimated modelWideResearchJobBenchSource
Ling 3.1 Flash83.3%59.1%estimated ± 5.1 pp, low confidence
Ornith-1.5-397B80.8%47.8%estimated ± 5.1 pp, low confidence
dots3-note Preview78.9%39.2%estimated ± 5.1 pp, low confidence
Qwen3.6 Plus74.3%18.5%estimated ± 5.1 pp, low confidence
Qwen3.5 397B74.0%17.2%estimated ± 5.1 pp, low confidence
Ling 3.0 Flash73.6%15.4%estimated ± 5.1 pp, low confidence
GLM-569.8%0.0%estimated ± 5.1 pp, low confidence
Ornith-1.5-35B-A3B67.8%0.0%estimated ± 5.1 pp, low confidence

Fitted on

ModelWideResearchJobBench
Atria Dawn Preview81.9%50.3%
Claude Opus 4.576.4%32.3%
Hy4 preview83.9%61.7%
Kimi K2.572.7%8.7%
Qwen3.8 Max81.9%53.4%