Anthropic’s AI improves 10 alignment benchmarks in 6 hours for $4/hr

Anthropic's automated alignment researcher (AAR) improved all 10 targeted misalignment benchmarks within six hours at a cost of about $4 per hour of inference, outperforming human researchers who cost $150 per hour. The system's best methods beat experienced human proposals, drastically reducing the time and cost of alignment training while potentially displacing human researchers in specific tasks.

TechCrunch https://techcrunch.com/2026/08/28/an-anthropic-researcher-just-gave-us-a-peek-at-self-improving-ai/ Anthropic’s AI improves 10 alignment benchmarks in 6 hours for $4/hr Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Anthropic’s automated alignment researcher AAR outperforms human researchers in improving AI alignment benchmarks at a cost of $4/hour versus $150/hour for humans, achieving better results within six hours. This drastically reduces the time and cost of alignment training, enabling rapid scaling of AI improvements while potentially displacing human researchers in specific tasks. Anthropic’s automated alignment researcher improved all 10 targeted misalignment benchmarks without degrading overall performance, and its best methods beat experienced human proposals within six hours at about $4/hour of inference versus $150/hour for human researchers. For production teams, the near-term leverage is automated post-training against well-defined evals, but the bottleneck shifts hard to benchmark quality: if your evals are incomplete or gameable, the system will optimize the wrong thing faster and cheaper than humans.