Claude now leads 26% of Anthropic’s AI research and development work, a sharp increase from less than 1% in February, as AI systems take on a growing role in building their successors.
Anthropic disclosed the figure in a new set of measurements designed to give outsiders more visibility into how quickly work inside frontier AI labs is becoming automated.
More than 90% of Anthropic’s measured AI R&D work now involves Claude at a level where the AI either collaborates with humans or leads the task, according to the company. Anthropic said Claude is not yet operating fully autonomously in any of the AI R&D work it measured.
The distinction matters because AI companies increasingly use their own models to help develop the next generation of AI. If that process becomes substantially automated, model development could accelerate while becoming harder for humans and outside observers to track.
Anthropic said its most-used internal platform had about 30,000 AI agents doing research and engineering work at any one time as of August. Those agents made more than 1 billion decisions during the month.
The company said all actions on that platform pass through automated monitoring. About 0.002% of decisions, or roughly one in 47,000, were blocked by an online monitor. An offline monitoring system flags roughly 100,000 transcripts a week, with about 50 of the highest-priority cases escalated to humans.
Anthropic is proposing three measurements that frontier AI developers could disclose publicly: how much AI R&D is performed by AI, how effectively companies monitor the actions of AI agents and how computing resources are allocated.
The company also disclosed that about 6% of the computing power devoted to AI R&D during one week in July went to safety work. Among computing resources used specifically for AI-driven AI R&D, about 12% went to safety.
Anthropic said the measurements could eventually be independently verified and compared across AI developers. The company plans to give outside evaluators access to internal processes, systems and data similar to the access provided to its own risk assessment teams.
The numbers provide an unusual view into a question that has largely been hidden inside AI labs: how much AI is already helping build AI.