Artificial intelligence is increasingly being used to develop the next generation of AI. On Thursday, Anthropic released a new set of metrics designed to track research and development progress at frontier AI labs. The data shows that, as of August, Claude has already taken the lead on roughly 26% of Anthropic's AI R&D work. Meanwhile, about 30,000 AI agents are simultaneously conducting research and engineering tasks within the company.

Still, Anthropic's data also reveals that AI remains far from fully autonomous AI development. As of August, Claude has not yet achieved fully autonomous operation in any of the AI R&D tasks measured. Anthropic says it hopes that by continuously disclosing these metrics, it can narrow the gap between the actual progress inside frontier AI labs and the information available to the public.

Claude Now Leads 26% of AI R&D Work

According to Anthropic's disclosed data, approximately 26% of the AI R&D work involving Claude has reached the "AI leads" level of automation as of August. This metric references Epoch AI's automation grading system for AI R&D. In this context, "AI leads" means that humans can provide higher-level instructions while the AI completes most of the specific work, with humans overseeing the process. By this standard, Claude's share of leading R&D work was less than 1% back in February. Currently, over 90% of related tasks have reached at least the "AI collaborates" level of automation.

However, Anthropic also emphasized that Claude has yet to reach full autonomy in any of the measured AI R&D tasks. This indicates that even though AI is deeply involved in model development and engineering processes, humans still remain a critical part of the R&D workflow.

30,000 Agents Working on Research and Engineering Simultaneously

Another metric disclosed by Anthropic shows that in August, roughly 30,000 AI agents were carrying out research and engineering work inside the company simultaneously. This scale reflects that AI agents are progressively evolving from simple assistive tools into actual productivity drivers within corporate R&D processes.

Anthropic also analyzed over 1 billion AI agent decisions made during August. The data shows that its online monitoring systems blocked approximately 0.002% of agent decisions, which translates to roughly one blocked decision per 47,000 decisions. This metric measures the proportion of decisions intercepted by monitoring systems, and it does not equate to the AI's error rate or the rate of dangerous behavior.

More Automated AI R&D Raises Growing Safety Concerns

Anthropic also shared details on how AI R&D computing resources are allocated. During the week spanning July 13 through July 20, roughly 6% of the company's AI R&D computing power was dedicated to safety work. When looking only at AI-driven AI R&D efforts, about 12% of computing resources went to safety-related tasks. As AI becomes increasingly involved in developing AI models themselves, AI safety becomes an ever more important issue.

Anthropic said it plans to give independent third-party evaluators access to its internal processes and data. This would allow for verification of safety practices and help outsiders understand the real progress of AI R&D inside frontier AI labs. This arrangement signals that AI laboratories are attempting to reduce the gap between internal corporate information and external public awareness through third-party assessments.

Moving from "AI-Assisted R&D" to "AI-Driven R&D"

The data Anthropic disclosed this time offers a new perspective on the growth of the AI industry. In the past, the market measured AI progress primarily through model parameters, benchmark scores, and user numbers. Anthropic is now trying to gauge the extent to which AI actually participates in developing the next generation of AI. Based on current data, AI has deeply integrated into the R&D processes of frontier AI labs, with vast numbers of agents handling research and engineering tasks, but a fully AI-driven end-to-end R&D process has not yet been achieved.

This also means that the transformation currently underway in the AI industry is less about "AI being able to independently develop AI" and more about AI rapidly increasing its level of involvement in AI research. As this share continues to grow, how to measure R&D automation levels, how to supervise AI agent behavior, and how to ensure safety investment keeps pace with the speed of R&D automation are likely to become ongoing questions that frontier AI labs will need to address.