Skip to main content

Anthropic's Claude Now Handles 26% of Its R&D Work

Anthropic's Claude Now Handles 26% of Its R&D Work

Anthropic has quietly released internal numbers that reveal just how deeply its own AI has burrowed into the company's research and development. As of August, Claude led 26% of all AI R&D tasks—a staggering leap from less than 1% just six months earlier.

Six Levels of Automation, But No Full Autonomy

The company uses a six-tier framework to track AI involvement. The fourth level, AL4, is defined as "AI-led": humans set the high-level goal, Claude does most of the heavy lifting, and a person supervises and reviews. Today, over 90% of R&D work hits that human-machine collaboration level or higher. Yet no task has reached AL5—full autonomy. AI always stays under human watch.

Behind the scenes, roughly 30,000 AI agents powered by Claude are working in parallel. They assign tasks to each other, cross-correct mistakes, and in August alone generated over 1 billion decisions in a single month.

80% of Code Comes from Claude

To keep risks in check, every AI operation gets an online safety check before execution. Abnormal decisions are intercepted outright. The background also samples conversation logs regularly, and high-risk cases go to human reviewers.

At the code level, the shift is even more obvious: over 80% of code merged into the repository is now generated by Claude. Engineers say their productivity has multiplied. These AI agents handle a massive amount of low-level work—model experiments, code writing, data analysis, and result verification. They even participate in iterating the next-generation Claude model, a process the industry calls recursive self-improvement.

Key Points

  • Claude's share of Anthropic R&D tasks jumped from <1% in February to 26% in August.
  • Over 90% of R&D work reaches human-machine collaboration, but no task is fully autonomous.
  • About 30,000 Claude-powered AI agents generated 1 billion decisions in August alone.
  • More than 80% of code merged into the repository is now written by Claude.
  • AI agents are involved in building the next Claude model—a form of recursive self-improvement.