Anthropic reports 80% of codebase written by Claude, highlights recursive self-improvement (xataka.com)

• Over 80% of Anthropic code integrated in May 2026 was Claude-generated
• Engineer productivity increased 8x per quarter
• Rekursive self-improvement enables autonomous AI training

"Anthropic revealed that in May 2026, over 80% of code integrated into their codebase was generated by Claude. Engineers now produce eight times more lines per quarter by reviewing AI output rather than writing code. The company describes recursive self-improvement, where AI models correct errors and train themselves continuously, leading to exponential capability growth but also raising control and safety concerns. Anthropic uses isolated environments and independent evaluators to mitigate risks."