Anthropic has disclosed that its Claude AI model has reached the point of "leading" 26 percent of the research and development work behind its own successor model. The disclosure is part of a broader effort to quantify how much AI itself is driving the development of future AI, alongside a new set of metrics designed to track the pace of that development.

Claude Now Leads a Quarter of R&D Work

Anthropic has introduced a six-level framework, AL0 through AL5, to classify how much AI is involved in internal R&D tasks. Work done entirely by humans sits at AL0, work where AI handles large chunks of a task under close human direction is classified as "collaborates," and work where AI can complete most of a task end-to-end from a high-level prompt is classified as "leads." The top tier, AL5, or full autonomy, has not yet been reached inside the company.

Under this framework, as of August 2026, Claude was found to be "leading" 26 percent of Anthropic's R&D work, up from under 1 percent in February 2026. More broadly, over 90 percent of the company's total R&D work now reaches the "collaborates" level or higher. Anthropic emphasized that Claude is not operating fully autonomously in any measured category of work, and that human oversight remains in place throughout.

Around 30,000 Agents Under Constant Monitoring

As of August 2026, roughly 30,000 AI agents were carrying out research and engineering work at Anthropic simultaneously. The company says it applies both "online" monitoring, which reviews agent actions in real time, and "offline" monitoring, which reviews records after the fact, across 100 percent of that agent activity. Of more than a billion decisions reviewed through online monitoring, only 0.002 percent were blocked. Offline monitoring flags roughly one to two transcripts per thousand for further review, with about 50 transcripts escalated to human reviewers each week.

Compute Spending on Safety Also Disclosed

Anthropic also published figures on how much of its computing resources are directed toward safety research. For the week of July 13 to 20, 2026, about 6 percent of the compute used for AI R&D overall went toward safety-related work, rising to about 12 percent when looking specifically at work where AI is driving its own R&D. The company defines a project as "safety-related" when its primary purpose is to make AI systems safer, more understandable, or more secure.

A Response to Industry-Wide Transparency Concerns

The disclosure comes as concern grows across the AI industry about how difficult it may become for humans to understand or control systems that are increasingly involved in accelerating their own development. Anthropic said it hopes that publishing measurements like these on an ongoing basis will make it easier for the public to track how AI development is actually progressing inside frontier labs, rather than relying on anecdotal or incomplete information. The company indicated it intends to keep updating these figures on a regular basis going forward.

Summary

Anthropic revealed that Claude now "leads" 26 percent of the R&D work behind its own next-generation model, alongside details of a roughly 30,000-agent monitoring system and figures on how much computing power is being spent on safety research. As debate continues over transparency in AI development, it remains to be seen how widely this kind of quantitative reporting will be adopted across the industry.