Anthropic has disclosed that its Claude AI models now lead approximately 26% of the company’s internal artificial intelligence research and development work — a dramatic increase from under 1% just six months ago in February 2026. The data comes from a newly developed internal tool called the R&D Automation Index, which tracks how much AI contributes across different research tasks.
The index uses a scale built by Epoch AI, ranging from AL0 (no AI involvement) to AL5 (fully autonomous operation). Claude currently operates at the AL4 level for 26% of the work, meaning the AI takes on most of a task after receiving a high-level instruction while a human remains in the loop to supervise. Anthropic emphasized that Claude is not yet running any operations fully on its own.
More than 90% of the company’s measured AI research now occurs at or above the level where AI and humans collaborate side by side. These areas include model training, reinforcement learning, evaluation, and engineering.
The scale of the operation is significant. In August, roughly 30,000 AI agents were active at any given time on Anthropic’s most-used internal platform. Over one billion agent decisions were tracked that month, of which approximately 0.002% were stopped by the company’s monitoring system. Additionally, about 6% of AI research compute was allocated to safety work during a week in July, rising to nearly 12% within AI-driven research specifically.
Anthropic plans to publish these measurements regularly and is calling on other leading AI developers to release similar data. The company believes that shared methods and external verification can help the broader industry keep pace with the rapid evolution of AI development itself. This report represents one of the most detailed looks any major AI company has offered into how its own tools are shaping the work of building the next generation of AI.