Anthropic says Claude now handles 25 percent of the work involved in developing its next-generation AI models

Anthropic has said its AI assistant Claude is now responsible for 26% of the artificial intelligence research and development work carried out within the company, highlighting how quickly AI systems are becoming involved in the creation of newer generations of AI technology. The San Francisco-based AI company said Claude also worked alongside human researchers on more than 90% of its research activities as of August. Anthropic announced the figures in a blog post as part of a new effort to regularly provide the public with data on how AI is being used in the development of increasingly advanced systems.

According to the company, Claude is still not working completely independently in any of the areas covered by the measurement. However, its role has expanded sharply. In March, Claude accounted for just 1% of the measured work, based on a framework developed by Epoch AI, an independent organisation that studies and tracks developments in artificial intelligence.

The figures provide an unusually detailed look at how a leading AI company is using its own technology to develop future models. They also come at a time when researchers and policymakers are increasingly examining the possibility that highly capable AI systems could eventually contribute to their own development with significantly less human involvement.

Greater AI autonomy has also raised questions about oversight and safety. Researchers have warned that as AI agents become more capable of acting independently, ensuring that their behaviour remains aligned with human intentions could become increasingly challenging. The issue has also prompted action from other major AI companies. OpenAI said on Wednesday that it plans to publish regular reports on unexpected or unauthorised AI behaviour, releasing six examples of concerning or unusual model activity as part of the initiative.

Anthropic said roughly 30,000 AI agents were carrying out research and engineering tasks on its main internal platform at any given time during August. The company said these agents operate under safeguards, with every action reviewed before it is executed. Out of more than one billion decisions made by the agents during the month, approximately one in every 47,000 decisions was blocked, according to Anthropic.

The company also provided details about the computing resources dedicated to AI safety. During a sample week in July, around 6% of Anthropic's computing capacity used for AI research was devoted to safety-related work. For research performed by AI systems themselves, that proportion increased to about 12%.

Anthropic noted that these numbers should be viewed as conservative estimates. Computing resources used for work that could simultaneously advance both AI capabilities and safety were classified as capability-related work rather than safety work. The disclosures offer a rare glimpse into the growing role AI systems are playing behind the scenes in the development of the next generation of artificial intelligence, while also underscoring the continuing importance of human oversight as these systems become more capable.