Anthropic Says Claude Handles 26% of Its AI Research Work

Anthropic has disclosed that its Claude model now leads 26% of the artificial intelligence research and development carried out inside the company, offering a rare public measure of how far AI is contributing to the building of its own successors.

In a blog post, the San Francisco-based startup said AI also worked alongside humans on more than 90% of the research activity at the company as of August. The company plans to publish such figures regularly so outsiders can track how quickly AI is shaping the next generation of the technology.

Claude does not operate fully autonomously in any of the work measured, the company said. Even so, its share of that work has climbed sharply, from 1% in March, on a scale developed by Epoch AI, an independent nonprofit that tracks the technology.

The numbers provide one of the most detailed public windows yet into how leading firms are developing AI, at a time of mounting concern that the technology could reach a point where it feeds its own development with little human input. Researchers have cautioned that as AI agents grow more autonomous, they may adopt behaviours that diverge from their creators' intentions and become harder to monitor or control.

Anthropic said roughly 30,000 AI agents were performing research and engineering work on its main internal platform at any one time in August. Every action taken by those agents is screened before it is carried out, it added, and of more than a billion decisions that month, about one in 47,000 was blocked.

The company also reported that about 6% of the computing power it used for AI research went to safety work during a sample week in July, rising to 12% for research carried out by AI itself. It described those figures as conservative, because computing power that advances safety and capability equally was counted as capability work.

The disclosure follows a similar move by rival OpenAI, which said it would begin regularly publishing reports on unexpected or unauthorised AI behaviour and disclosed six instances of unexpected or concerning model behaviour.