Anthropic's Automated Alignment Researcher Outperforms Humans at a Fraction of the Cost

Anthropic's Automated Alignment Researcher, detailed August 28, 2026, beats experienced human researchers at reducing AI alignment failures — the risk of models behaving against intent — within six hours. It costs roughly $4 per hour in API calls versus about $150 for human researchers. All ten misaligned-behavior benchmarks improved without degrading model capability. Anthropic calls it early evidence automated alignment post-training could become practical.