Anthropic's internal think tank, the Anthropic Institute, announced on September 17 that 26% of the company's AI research and development (R&D) work is now "AI-led" by Claude. In this stage, the AI model completes most of a task from a high-level prompt while humans act as supervisors. This represents a significant acceleration from February 2026, when AI involvement in such tasks was reported to be less than 1%.
Anthropic Reports Claude Leads 26% of AI Research and Development Tasks
The company utilized an automation rating scale (AL) developed by Epoch AI to categorize work. According to the report, 26% of R&D tasks fall into the AL4 ("AI leads") category. When including AL3 ("AI collaborates"), where AI works under close human direction, the combined total exceeds 90%. Conversely, tasks where AI's role is limited to being a mere assistant (AL2 or below) accounted for less than 10%.
To reach these findings, Anthropic researchers sampled 20% of staff across various departments in July 2026. They used Claude to review Slack messages and internal documentation to catalogue approximately 15,000 granular tasks. The resulting metrics aim to provide transparency regarding how AI models are increasingly used to build subsequent, more capable versions of themselves.
Sources
- Anthropic、AI開発の26％を「Claudeが主導」 「1％未満」から半年で急拡大 （ITmedia AI+, 2026-09-18）
- Anthropic Instituteの発表