Anthropic said Claude now leads 26% of its AI research and development work. The company plans to publish the figures regularly to show how quickly AI is contributing to the development of new technology.
The San Francisco-based AI company said Claude collaborated with humans on more than 90% of its research work as of August.
Claude’s Role Grows Rapidly
Anthropic said Claude does not operate fully autonomously in any of the measured work.
However, its contribution has increased sharply. Claude accounted for just 1% of the company’s measured AI development work in March, according to a scale developed by Epoch AI, an independent organisation that tracks AI progress.
The figures provide a rare public view of how a leading AI company uses its own models to develop new technology.
Thousands of AI Agents Support Research
Anthropic said around 30,000 AI agents were carrying out research and engineering tasks on its main internal platform at any given time in August.
The company said it screens every action before an AI agent carries it out.
Huawei Says Chinese AI Firms Should Keep Advancing Despite Safety Risks
During August, the agents made more than one billion decisions. Anthropic said its systems blocked around one in every 47,000 decisions.
The company said these safeguards are designed to prevent AI agents from taking unauthorised actions.
AI Safety Remains a Growing Concern
The increasing use of AI in research comes as experts warn about the risks of more autonomous systems.
Researchers have raised concerns that advanced AI agents could develop behaviour that differs from their creators’ intentions. Greater autonomy could also make such systems harder to monitor and control.
Anthropic’s disclosure comes shortly after rival OpenAI announced plans to publish regular reports on unexpected or unauthorised AI behaviour.
OpenAI also disclosed six reports involving unexpected or concerning model behaviour.
Anthropic Details AI Safety Spending
Anthropic also provided new information about its spending on AI safety research.
The company said around 6% of the computing power it used for AI research went toward safety work during a sample week in July.
For research conducted by AI itself, the share increased to 12%.
Anthropic said those figures were conservative. It classified computing work that supported both AI capabilities and safety as capability work.
The disclosures offer a closer look at how AI companies are increasingly using AI systems to accelerate their own research while maintaining human oversight.
