Claude steers more than a quarter of Anthropic's artificial-intelligence research and development work, said the frontier lab as it sought to offer better visibility on its progress.

AI assistant Claude led 26% of Anthropic's AI R&D as of August, up from effectively zero percent a year earlier, the company said in a report Thursday.

"AI systems are getting more powerful, and they're increasingly being used to build the next version of themselves," said Anthropic. "We want to illuminate the pace of progress for the public."

The company measured this through a prototype index that catalogues every type of AI R&D work done at the company, rating how automated each task is and aggregating those ratings.

The report found that the share of work where Claude collaborates closely with staff is around 90%, said the company.

While AI R&D is driven by Claude, the AI assistant isn't operating fully autonomously for any measured subset of this work, said Anthropic.

This measurement is among a series built by Anthropic that can shed light on critical aspects of AI development, such as the extent to which AI is building the next version of itself and the company's ability to oversee and intervene in actions that AI agents take, said Anthropic.

It plans to embed independent third-party evaluators from multiple organizations at Anthropic, and give them access to data "comparable to what internal risk assessment teams have," it said.

Anthropic said it created a system that allows it to oversee actions that AI agents take on the company's systems. As of August, approximately 30,000 AI agents did research and engineering work at Anthropic's most-used internal platform at any one time.

The company also developed a framework to measure its compute allocation to gauge the amount of resources that go toward safety testing and evaluation. The company found that about 6% to 12% of its computing resources for AI-related R&D went toward safety in a week in July.

Concerns over AI's existential risks have risen recently, with some researchers leaving key AI companies, citing fears that the rapidly advancing technology could eventually destroy humanity.

In a rare display of unity among leaders of AI companies, OpenAI's Sam Altman, Anthropic's Dario Amodei, and Elon Musk, who founded xAI, warned that they needed to slow development of the technology before these advances create something that can't be controlled.

OpenAI on Wednesday unveiled a framework for reporting instances of misbehavior by its AI models and disclosed six new examples of such misbehaviour.

News Corp, owner of The Wall Street Journal and Dow Jones Newswires, has a content-licensing partnership with OpenAI.