Anthropic published a research report on June 4, 2026, revealing that AI systems now write the overwhelming majority of code at the company. By May 2026, more than 80% of code merged into Anthropic's production codebase was authored by Claude, up from low single-digit levels in February 2025. The report, titled "When AI Builds Itself: Our progress toward recursive self-improvement," details how AI is increasingly taking over AI development itself.
Engineers Merge 8x More Code With AI Assistance
The productivity gains from AI-assisted development have been dramatic:
- Engineers now merge approximately 8 times more code per quarter compared to 2024 levels
- In March 2026, Anthropic researchers reported producing roughly 4 times as much output using advanced AI models compared to working without such assistance
- Claude Opus 3 handled 4-minute tasks in 2024; by 2026, Claude Opus 4.6 managed 12-hour tasks
- In April 2026, Claude shipped over 800 fixes that reduced a class of API errors by a factor of one thousand, with the supervising engineer estimating a human would have taken four years to complete this work
Claude's Success Rate on Open-Ended Problems Reaches 76%
Anthropic measured progress across two dimensions: execution capability and direction-setting judgment. Claude's success rates on open-ended problems reached 76% in May 2026. More significantly, Claude's ability to choose research directions improved from beating human suggestions 51% of the time in November 2025 to 64% in April 2026.
The research separates execution (running experiments, writing code) from strategic judgment (choosing which problems matter). If AI eventually masters both, development pace would be determined solely by computational resources rather than human labor.
Company Calls for Potential Pause on AI Development
Anthropic's head of internal research Marina Favaro and head of policy Jack Clark argued that a pause would provide breathing room. Clark suggested some models could be capable of recursive self-improvement within just two years. The company emphasized: "We are not there yet, and recursive self-improvement is not inevitable, but it could come sooner than most institutions are prepared for."
The Hacker News discussion garnered 313 points and 414 comments as of June 4, showing significant community debate about the implications of AI-driven AI development and whether the claims represent genuine progress toward recursive self-improvement or sophisticated tooling.
Key Takeaways
- By May 2026, more than 80% of code merged into Anthropic's production codebase was authored by Claude, up from low single-digit levels in February 2025
- Engineers now merge approximately 8 times more code per quarter compared to 2024 levels
- Claude's success rate on open-ended problems reached 76% in May 2026, while its ability to choose research directions improved from 51% to 64%
- In April 2026, Claude completed 800 fixes in a timeframe that would have taken a human engineer four years
- Anthropic's policy head Jack Clark suggested some models could be capable of recursive self-improvement within just two years