According to data disclosed by Anthropic, Claude already leads 26% of internal AI research and development efforts, collaborating on more than 90% of related tasks. "Leading" means that humans provide high-level instructions, while the AI handles most of the work, with humans taking responsibility for oversight. At the same time, Claude has not yet achieved fully autonomous operation in any of the AI projects it has been tested on.
AI is increasingly involved in developing the next generation of AI.
On Thursday, local time, Anthropic unveiled a new set of metrics designed to track the R&D progress of cutting-edge AI labs. The data show that, as of August this year, Claude has "led" roughly 26% of Anthropic's AI research and development efforts; meanwhile, approximately 30,000 AI agents are simultaneously engaged in both research and engineering at the company.
However, Anthropic's data also indicate that AI remains far from achieving full autonomy in conducting AI research and development. As of August, Claude had not yet attained a level of fully autonomous operation in any of the AI R&D tasks tested.
Anthropic stated that it hopes to narrow the gap between the actual progress made within leading AI labs and the information available to the public by consistently disclosing these metrics.
Claude has led 26% of AI research and development efforts.
According to data disclosed by Anthropic, as of August, approximately 26% of the AI research and development projects involving Claude had already reached an "AI‑led" level of automation.
This metric draws on Epoch AI's AI‑driven R&D automation rating framework. Under this framework, "AI‑driven" signifies that humans provide high‑level directives while AI handles the bulk of the execution, with human oversight remaining in place. By this measure, Claude‑driven R&D accounted for less than 1% as of February this year.
Currently, more than 90% of relevant tasks have reached an automation level of "AI collaboration" or higher.
However, Anthropic also emphasized that Claude has not yet achieved full autonomy in any of the AI research and development projects it has evaluated. This means that, despite AI's deep integration into model development and engineering workflows, humans remain a critical component of the R&D process.
30,000 agents are simultaneously conducting research and engineering work.
Another metric disclosed by Anthropic indicates that in August this year, approximately 30,000 AI agents were simultaneously engaged in research and engineering tasks within the company.
This scale reflects that AI agents are evolving from mere auxiliary tools into genuine productivity drivers within enterprise R&D workflows.
Anthropic also analyzed more than one billion AI agent decisions made in August. The data show that the online monitoring system intercepted approximately 0.002% of these decisions—roughly one out of every 47,000 decisions.
This metric measures the proportion of decisions blocked by the monitored system and is not equivalent to the AI's error rate or the incidence of risky behavior.
The more automated AI research and development becomes, the greater the focus on security concerns.
Anthropic also disclosed the allocation of computing resources for its AI research and development.
During the week of July 13 to July 20, approximately 6% of the company's AI research and development computing resources were allocated to security-related tasks; when considering only AI‑driven R&D efforts, about 12% of these resources were devoted to security‑related work.
As AI increasingly plays a role in the development of AI models themselves, AI safety has become an ever-more critical issue.
Anthropic stated that the company plans to grant independent third-party assessment organizations access to its internal processes and data, in order to verify its safety practices and help external stakeholders gain insight into the actual progress of AI research and development within cutting-edge AI labs.
This arrangement signifies that AI labs are seeking to narrow the gap between internal corporate information and public perception through third-party assessments.
From "AI-Assisted R&D" to "AI-Driven R&D"
The data disclosed by Anthropic offers a fresh perspective on the evolution of the AI industry: whereas the market has traditionally gauged AI progress primarily through model parameters, benchmark scores, and user numbers, Anthropic is now seeking to assess the extent to which AI is actively shaping the development of the next generation.
Based on current data, AI has already been deeply integrated into the R&D workflows of cutting-edge AI laboratories, with numerous agents taking on both research and engineering tasks. However, fully autonomous completion of the entire R&D process by AI remains unattainable.
This also means that the changes currently underway in the AI industry are less about "AI being able to independently develop AI" and more about AI rapidly increasing its own involvement in AI research and development.
As this ratio continues to rise, questions such as how to measure the level of R&D automation, how to monitor the behavior of AI agents, and how to ensure that security investments keep pace with the speed of R&D automation may become ongoing challenges for cutting-edge AI labs.
Editor/lambor