When AI Builds Itself: Anthropic's progress toward recursive self-improvement

What happened

Anthropic released a position paper from its Institute arguing AI systems are increasingly automating AI development itself, with task-length doubling every four months.

Context and impact

The company says it is not yet at full recursive self-improvement, but the milestone could arrive sooner than most institutions are prepared for. It proposes international coordination and verification mechanisms enabling a frontier pause if needed.

Details

• 80%+ of merged production code at Anthropic was authored by Claude (May 2026)
• Claude reached a 76% success rate on open-ended programming tasks (+50 pp in six months)
• Task-completion doubling time dropped from 7 to 4 months
• ~52x speedup on code-optimization tasks (April 2026) vs. ~3x in May 2025
• Automated Claude review would catch ~one-third of production bugs pre-deploy