Anthropic says its Claude AI is now playing a major role in developing the company’s next generation of models — a milestone that is intensifying the debate over how quickly artificial intelligence should be allowed to improve itself.
The AI company said Claude now leads about 26% of its artificial-intelligence research and development work, meaning the model can handle large portions of a project from a high-level instruction through completion, while remaining under human supervision.
That figure was zero in February, according to Anthropic. By August, Claude was leading more than a quarter of the company’s R&D work. Overall, Anthropic said roughly 90% of its AI research and development now involves Claude working alongside human researchers.
Claude Is Helping Build the Next Claude
Anthropic’s disclosure represents a significant shift in the way frontier AI systems are developed.
Rather than being used only as a consumer chatbot or coding assistant, Claude is increasingly being deployed inside Anthropic’s own research process to help with tasks involved in creating more advanced AI models.
Anthropic said its engineers are delegating an increasing share of AI-development work to AI systems themselves, accelerating the development cycle.
The company describes the longer-term possibility as “recursive self-improvement” — a scenario in which an AI system becomes capable of autonomously designing and developing its own successor.
But Anthropic emphasized that this has not happened yet. Human researchers remain involved, and Claude is not currently operating as a completely autonomous system that designs and builds the next model without people.
From Zero to 26% in Just Months
The speed of the change is what has attracted particular attention.
Anthropic said Claude’s share of R&D work that it “leads” rose from 0% in February to 26% in August 2026.
The company also reported that approximately 30,000 AI agents were performing research and engineering work under oversight mechanisms as of August.
Anthropic says these systems can complete substantial portions of technical work, but monitoring remains necessary because increasingly capable AI agents can behave in unexpected ways.
The Washington Post similarly described the 26% figure as an indicator of progress toward AI systems capable of contributing to their own development.
Why “AI Building AI” Matters
The development has implications beyond Anthropic.
If AI systems can increasingly perform the research, coding, testing and evaluation required to create more capable models, companies could potentially shorten the time needed to develop each new generation.
That creates a feedback loop: better AI helps researchers build better AI, which could then help build an even more capable generation.
Anthropic’s own research describes this as a trend toward recursive self-improvement, while stressing that full autonomous self-improvement is neither present nor inevitable.
The company has also been researching ways to automate AI-safety work. In an August study, Anthropic reported experiments in which Claude autonomously trained models to improve their performance on benchmarks measuring different alignment failures.
Safety Concerns Are Growing Alongside Capability
The announcement comes at a particularly sensitive moment for the AI industry.
Anthropic CEO Dario Amodei has recently called for greater coordination and a slower pace of frontier AI development, arguing that companies need more time to establish safeguards as systems become increasingly capable.
Other technology leaders have also been debating how quickly the industry should advance and what safeguards should accompany increasingly autonomous AI agents.
Anthropic itself has argued that society needs better measurements of how rapidly AI capabilities are progressing, particularly because the public may otherwise have little visibility into how much AI is contributing to the development of future systems.
The Bigger Question: Who Is Building Whom?
Anthropic’s latest figures do not mean Claude has become an independent AI researcher or that humans have lost control of model development.
Instead, they show that AI systems are becoming increasingly embedded in the research-and-development pipeline used to create the next generation of AI.
That distinction matters.
Today, Claude is working under human direction. But the more development work AI systems can perform themselves, the more important it becomes to understand where human oversight remains essential — and whether existing safety measures can keep pace with rapidly increasing capability.
Anthropic says it wants other AI developers to publish comparable measurements so the public can better understand how close the industry may be to genuine recursive self-improvement.
For now, the headline is not that AI has completely learned to build itself.
It is that AI is increasingly helping humans build the next generation of AI — and that transition is happening faster than it was only months ago.
