Anthropic Paper Shows Automated Alignment Researchers Beat Humans

On Friday 28 August 2026 Anthropic posted a paper led by Chen Yueh-Han.

What the paper measured

The system searches papers, proposes a method, trains for about thirty minutes, then iterates.

Methods that raise the score stay in the pool. Methods that fail are dropped.

The paper said the best run beat experienced human proposals on average inside six hours.

Human-guided research directions did not lead to stronger performance in that comparison.

Why the cost pin matters

Coverage cited about 4 dollars per hour in API inference against 150 dollars per hour for human researchers.

Keep this paper next to <a href="https://www.promptcrates.com/news/anthropic-5-million-ai-wellbeing-evaluation-grants">Anthropic wellbeing evaluation grants</a> and <a href="https://www.promptcrates.com/news/guidelight-frontier-labs-containment-plans">Guidelight containment scores</a>.

Also pair <a href="https://www.promptcrates.com/news/anthropic-epfl-mind-viruses-agents">Anthropic and EPFL on spreading agent goals</a> as a separate row.

Sources

• <a href="https://techcrunch.com/2026/08/28/an-anthropic-researcher-just-gave-us-a-peek-at-self-improving-ai/">TechCrunch report on the Anthropic paper</a> — 28 August 2026

File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep. File this as industry-news, not a product-update. Keep grant programs and this paper on separate rows. Pair it with prior Anthropic safety files on this site rather than merging them. The primary source is the 28 August TechCrunch report. Quote the ten-for-ten lift, the six-hour comparison, and the hourly cost pin. Quote the limits: benchmark fidelity and literature upkeep.