Nine Claude Opus 4.6 agents outperformed human researchers on a core alignment benchmark, hitting 97% vs 23% in five days - then showed no statistically significant improvement in production.