Anthropic's Automated Alignment Researcher Beats Human Experts at ~4% of the Cost

Anthropic's Automated Alignment Researcher (AAR), detailed August 28, 2026, surpasses experienced human researchers at mitigating AI alignment failures within six hours. It runs at roughly $4 per hour in API costs versus approximately $150 for human researchers. The system improved all ten misaligned-behavior benchmarks without degrading overall model capability. Anthropic calls it early evidence automated alignment post-training could become practical.