Engineers at Anthropic are producing eight times more code in 2026 than they did between 2021 and 2025. The reason is not that they are working longer hours—it is that artificial intelligence now generates most of the code. For the first time, the company has published internal data showing how AI is accelerating the development of AI itself.
The numbers are striking. In February 2025, the share of code written by Claude amounted to only a few percentage points. By May 2026, that figure had surpassed 80% of all code being incorporated into Anthropic’s codebase. Engineers still guide the process, define objectives, and review outputs, but they no longer spend most of their time manually writing code.
A similar trend is emerging in AI research. In April 2026, Anthropic conducted an experiment in which AI agents were assigned an open-ended model safety problem and allowed to work independently. The agents generated hypotheses, designed experiments, shared findings with parallel agents, and iteratively refined their approaches. According to the company, two human researchers completed approximately 23% of the task within one week. The AI agents, after accumulating roughly 800 hours of combined work, completed 97% of the same challenge.
Public benchmark results show a comparable rate of progress. In May 2026, Claude achieved a 76% success rate on tasks without clearly defined solutions—problems where the correct answer is not immediately obvious. Anthropic notes that this represents an improvement of roughly 50% compared with performance six months earlier.
The company describes these developments as a direct path toward recursive self-improvement—a scenario in which an AI system becomes capable of designing and training its own successor. According to Anthropic, that stage has not yet been reached. Humans still determine which problems should be addressed and remain responsible for evaluating outcomes. However, the gap between what people do and what AI systems can accomplish is narrowing rapidly.
Anthropic cautions that such advances create not only opportunities for scientific research, medicine, and technology, but also genuine risks related to maintaining human oversight and control. In response, the company has announced the creation of the Anthropic Institute, an organization dedicated to studying mechanisms for slowing or temporarily pausing the development of frontier AI models—provided that other major AI developers adopt similar measures and that compliance can be independently verified.
Source: Anthropic
Image: AI-generated
