Claude Self-Improvement Takes a Step Forward as AI Handles 26% of Model R&D

Anthropic disclosed this week that Claude now “leads” 26% of the company’s model research and development in just seven months, up from virtually none earlier this year. Claude also collaborates with humans on more than 90% of Anthropic’s research work. 

The figures accompany a new Anthropic institute post titled “When AI builds itself,” which examines how AI is accelerating the development of newer AI systems. Anthropic frames this as an early step toward recursive self-improvement, although it stresses that Claude is not autonomously developing its successor.  

How Much of Anthropic's R&D Does Claude Now Run? 

“Leading” a task means Claude can receive a high-level goal and complete most of the work without detailed step-by-step instructions, while humans continue to supervise the process.  
 
“Collaborating,” the broader figure covering more than 90% of Anthropic’s research work, includes tasks in which Claude performs substantial portions of the work under closer human direction. 

The shift is also visible in Anthropic’s codebase. As of May 2026, more than 80% of the code merged into production was attributable to Claude, up from the low single digits before Claude Code launched in research preview in February 2025. 

In the second quarter of 2026, the typical Anthropic engineer merged approximately eight times as much code per day as in 2024. Anthropic cautions that lines of code measure output volume rather than productivity or quality. 

Why Recursive Self-Improvement Worries AI Researchers? 

Anthropic flags the risk in its own post, writing: “We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.” 
 
But the company also warns that a system accelerating its own development could get harder for humans to understand or steer as AI model autonomy grows. That concern echoes Microsoft AI CEO Mustafa Suleyman’s recent warning that Anthropic’s approach could make Claude harder to control.  

Anthropic says approximately 30,000 AI agents were performing research and engineering work on its main internal platform at any one time in August 2026. According to the company, every proposed agent action was screened before execution. Of more than one billion decisions reviewed that month, roughly one in 47,000 was blocked. 

A Disclosure Arriving Amid Calls to Slow AI Down 

The timing is not incidental. CEO Dario Amodei has been among the most vocal advocates for pacing frontier AI development. OpenAI’s Sam Altman and X’s Elon Musk have echoed the same in recent weeks. 

Days earlier, Anthropic research Jacob Coxon resigned, warning on social media that AI laboratories were “racing straight to self-improving superintelligence and gambling with our lives.”  

Anthropic says it plans to publish these development metrics regularly. It has also called for greater industry-wide transparency so that policymakers, researchers and the public can track the pace of AI development from outside the laboratories building the systems.