Anthropic says its Claude artificial intelligence models are now leading more than a quarter of the company’s own research and development work, offering a rare look at how quickly AI is becoming involved in building the technology itself.
The San Francisco-based AI company said Thursday that Claude “leads” 26% of the research and development work measured inside Anthropic, up sharply from 1% in March. At the same time, AI collaborated with humans on more than 90% of the research work conducted by the company as of August.
Anthropic said it plans to publish the measurements regularly, using a scale developed by Epoch AI, an independent nonprofit that tracks AI progress. The initiative is intended to provide outsiders with a clearer view of how much AI is contributing to the development of sophisticated systems.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
The figures arrive as researchers and AI companies grapple with a central question surrounding the next phase of the technology: Will AI systems increasingly be able to improve the tools and systems that produce them, reducing the amount of direct human involvement required?
Anthropic said Claude is not operating fully autonomously in any of the areas covered by its measurement. Still, the rapid increase in its contribution shows how quickly AI-assisted research has expanded within the company. The company said about 30,000 AI agents were performing research and engineering work on its main internal platform at any given time in August. Every action taken by those agents is screened before execution.
Of more than 1 billion decisions made by the agents during the month, roughly one in 47,000 was blocked by the company’s controls. That scale illustrates both the growing role of AI agents in technical work and the monitoring challenge that comes with deploying large numbers of them.
As AI systems take on more complex tasks, researchers have raised concerns that greater autonomy could produce behavior that diverges from the intentions of their developers, making systems more difficult to monitor or control.
Anthropic’s disclosure also provides a measure of the resources the company is directing toward AI safety as its models become more deeply embedded in its research process. The company said about 6% of its computing power devoted to AI research went toward safety work during a sample week in July. For research conducted by AI itself, the share was higher, at 12%.
Anthropic said those figures are conservative because computing used for work that simultaneously advances AI capabilities and safety is classified as capability work rather than safety work.
The disclosure comes as AI companies face growing pressure to provide more information about unexpected model behavior. OpenAI, another leading AI developer, said Wednesday that it would begin regularly publishing reports on unexpected or unauthorized behavior. The company also disclosed six incidents involving unexpected or concerning behavior by its models.
The moves by both companies point to an industry increasingly focused not only on how quickly models are improving, but also on measuring what those models are doing inside the companies developing them.
Anthropic’s data is notable because it provides a glimpse into a potential feedback loop in AI development. Claude is being used to perform research and engineering tasks that can contribute to the development of future AI systems, even though humans remain involved in the process.
More than 90% of Anthropic’s measured research work involved collaboration between AI and humans in August, suggesting that the company is currently operating a hybrid model rather than handing development entirely to autonomous systems.
Anthropic’s 26% figure refers to work that Claude “leads,” while the broader 90% figure covers research in which AI collaborated with humans. The company said neither category represents fully autonomous AI development. That leaves humans central to Anthropic’s research process, while the role of AI is expanding rapidly.
The trend also puts greater emphasis on the safeguards surrounding AI agents. With tens of thousands of agents carrying out technical tasks and more than a billion decisions processed in a single month, even a very low rate of blocked actions represents a large number of opportunities for automated systems to encounter activities that require intervention.
Anthropic’s decision to publish the data regularly could become a useful benchmark for tracking how much AI is actually contributing to AI development, rather than relying solely on claims about autonomous systems. The company’s latest figures show that AI has not yet taken humans out of the loop at Anthropic. But they also show how quickly that loop is changing, with Claude’s measured contribution to the company’s own research rising from 1% in March to 26% in August.