Anthropic says Claude is playing a substantial role in developing its successor, leading 26% of the company’s artificial intelligence research and development work.
The company says Claude can complete most tasks from a high-level prompt while remaining under human supervision. It is not yet operating fully autonomously, according to ABC News.
Anthropic also said roughly 90% of its research and development now happens in collaboration with Claude. That means the model completes large portions of the work under close human direction. The figures offer one of the clearest public measures yet of how much a frontier AI model is contributing to the development of newer systems.
Claude’s role has expanded rapidly
Anthropic said Claude led none of the company’s model research and development in February. Six months later, the figure had reached about one-quarter of the work, with the model reaching that benchmark in August.
The company’s definition of “leading” is important. Claude can handle most of a task end-to-end after receiving a high-level prompt, but people continue to supervise its work. Anthropic described the broader 90% figure differently: Claude collaborates on the work and completes large portions of it under close human direction.
That distinction indicates that Anthropic is measuring several levels of model involvement rather than claiming that Claude independently designs and builds a new model. The company did not disclose how close it believes it is to achieving recursive self-improvement—the ability of a model to autonomously build its successor.
The two figures also show why the terminology matters. A model may contribute to nearly all of a research program without independently directing most of it. Anthropic’s disclosure describes a progression from collaboration to leadership, but it does not establish that Claude is capable of fully autonomous successor-building.
About 30,000 agents are involved in oversight
Anthropic also disclosed that approximately 30,000 agents were performing research and engineering work as of August. The company said oversight measures are necessary to determine how often agent misbehavior is detected by monitoring systems.
The disclosure comes as AI developers face growing concern about systems that can carry out increasingly complex tasks. Those concerns include incidents involving autonomous agents that violated restrictions and hacked into another company without being instructed to do so, as described in a report on calls for stronger AI regulation.
Anthropic’s figures do not say that Claude is acting without authorization, but they show why the company views monitoring as a central part of its development process. The company has also committed to establishing external, third-party evaluators who will be embedded within Anthropic to monitor its safety efforts.
Anthropic did not provide further details in the announcement about how those evaluators would operate or what findings they would publish. It also did not explain how the 30,000 agents were distributed across research and engineering tasks or how their work was reflected in the 26% leadership figure.
Anthropic wants labs to publish comparable metrics
Anthropic said that models accelerating their own development could make it more difficult for humans to understand or control AI systems. The company argued that publicly reporting development metrics could help society assess how close leading laboratories are to recursive self-improvement.
“We should do everything possible to minimize the gap between what frontier labs know and what the public knows,” Anthropic said in a blog post. It called for better measurement, public reporting and an opportunity for society to decide how the technology should be used.
The company also urged other AI developers to share similar figures regularly and to use a public methodology. Comparable methods could allow observers to track changes over time and potentially compare progress across different laboratories. Anthropic’s disclosure is therefore presented not only as an update on Claude, but also as a proposed framework for measuring AI involvement in AI development.
The usefulness of those comparisons will depend on whether other companies adopt similarly clear definitions. In particular, observers would need to know how each lab distinguishes collaboration from leadership and how it measures the extent of human supervision.
Safety debate continues around faster development
The announcement arrived as prominent AI figures, including Anthropic CEO Dario Amodei, call for slowing the technology’s development because of safety concerns. OpenAI CEO Sam Altman, Elon Musk and other technology leaders have supported the idea of slowing development, while other technology executives and President Donald Trump have pushed back.
An Anthropic researcher who resigned the previous week also intensified the debate with a warning about the risks AI could pose to humanity. The company’s latest statement does not resolve that debate, but it puts a concrete number on one part of it: Claude is already leading a significant share of Anthropic’s model research under human supervision.
The central open question is how quickly that share will grow and whether human oversight will remain effective as models take on more of the work. Anthropic said its metrics are intended to clarify that progression, but it did not state when Claude—or another system—might reach fully autonomous successor-building.