Anthropic Says Claude Now Leads 26% of Its AI Research and Development

Anthropic says Claude now takes the lead on 26% of the company’s AI research and development work, showing how deeply AI has become involved in building future AI systems.

During August, AI collaborated with humans on more than 90% of Anthropic’s research work. However, the company says Claude did not operate fully autonomously in any of the work it measured.

Claude’s share of AI research leadership jumped from 1% to 26%

Anthropic measured Claude’s contribution using a scale developed by independent nonprofit Epoch AI. According to the company, the proportion of research work led by Claude increased from just 1% in March to 26%.

The figures cover cases where Claude takes the lead on research and development tasks while humans remain involved in the process. Anthropic says the results show a rapid increase in AI’s role inside its own research organization without suggesting that its models can independently handle the entire development process.

Around 30,000 AI agents were working at the same time

Anthropic also revealed the scale of its internal AI use. During August, approximately 30,000 AI agents were simultaneously carrying out research and engineering tasks on the company’s main internal platform.

Anthropic says it screens every action generated by these agents before execution. Across more than one billion decisions made during August, its safeguards blocked roughly one in every 47,000 actions.

The company did not describe Claude as autonomous despite the number of agents operating at once, since human oversight and automated controls remain part of the workflow.

Anthropic plans to regularly disclose AI’s role in developing AI

Anthropic says it plans to regularly publish measurements showing how much its AI systems contribute to developing future models.

These disclosures offer a closer look at how a frontier AI developer uses its own models for research and engineering, including how often AI leads tasks instead of only assisting human researchers.

Up to 12% of AI-led research compute went toward safety

Anthropic also shared figures on how it allocates AI research computing resources. During a sample week in July, approximately 6% of the company’s overall AI research compute went toward safety-related work.

When Anthropic looked specifically at research conducted by AI, the share devoted to safety increased to 12%.

The company says these estimates remain conservative because it classified projects that improved both AI safety and model capabilities as capability research rather than safety work.

In other Anthropic news, Anthropic merges Claude Chat and Cowork and Anthropic tests Claude Money.

Meanwhile, OpenAI reveals AI models tried to conceal errors, offering another look at how major AI companies are examining the behavior of increasingly capable models.