Claude now leads 26% of Anthropic’s next-generation technology work

Anthropic’s latest figures show how quickly AI is moving from assistant to research collaborator, while one important limit remains in place.

On September 17, Anthropic announced that its artificial intelligence system, Claude, now plays a leading role in 26% of the research and development work involved in creating next-generation technologies. The company plans to publish such figures regularly to demonstrate the pace of AI development.

As of August, AI was collaborating with humans on more than 90% of Anthropic’s research projects. At the same time, Claude was not yet operating fully autonomously in any of the areas assessed.

Claude’s contribution to technology development had grown from 1% in March. The assessment used a scale developed by the independent nonprofit organization Epoch AI, which tracks the progress of artificial intelligence.

How AI Participates in Internal Work

The published figures make it possible to assess in greater detail how leading companies in the industry are using AI for their own research and development. The data emerged amid concerns that these systems could approach a level at which they would be able to improve technology with limited human involvement.

Researchers warn that as AI agents become more autonomous, the risk increases that they will exhibit behavior that deviates from their developers’ intentions. This could also make it more difficult to monitor and control such systems.

In August, around 30,000 AI agents were operating simultaneously on Anthropic’s internal platform, carrying out research and engineering tasks. Every action taken by the agents was checked before it was executed.

Over the course of the month, the agents made more than one billion decisions, and approximately one in every 47,000 was blocked during review.

Computing Resources Devoted to Safety

The company also disclosed data on its use of computing power. During a selected week in July, about 6% of the resources Anthropic devoted to AI research went toward safety work.

For research conducted autonomously by AI, the figure was 12%. The company noted that these figures were conservative: computing resources used simultaneously to improve safety and expand model capabilities were counted as work to enhance functionality.

Separately, competitor OpenAI announced plans to publish regular reports on unexpected or unauthorized behavior by its systems. The company also disclosed six instances of unusual or potentially dangerous model behavior.

The growing share of work performed by AI systems demonstrates their increasingly important role in creating new models, while also underscoring the need for continuous human oversight.

We recommend paying attention to: