K-Bench preprint finds no tested scientific agent clears all judge thresholds

A new preprint reports K-Bench 01, an evaluation of nine frontier models on 1,602 completed scientific-agent runs drawn from first-turn requests sampled from K-Dense Web traffic. The authors say no model met their scientist-acceptance threshold under all three blinded language-model judges. They report gpt-5.6-sol had the highest pooled mean, while its confidence interval spanned the threshold and two judges ranked claude-opus-5 first. Across 39,934 scored judgments, 47.6% fell below the threshold.