AI
Anthropic's Claude Now Leads 26% of Its Own Development Work
The AI company disclosed that its model handles end-to-end tasks under supervision as the industry debates recursive self-improvement risks.
Key takeaways
- Claude now leads 26% of Anthropic's model R&D work, up from 0% in February, handling end-to-end tasks under human supervision.
- The AI model participates in roughly 90% of the company's research through collaboration, doing large chunks of work under human direction.
- Anthropic is calling on other AI labs to publicly share metrics on recursive self-improvement to help society understand and govern the technology.
- The company operates approximately 30,000 agents for research and engineering work with monitoring systems to detect misbehavior.
- The disclosure comes amid industry debate over AI safety, with tech leaders divided on whether to slow development.
Anthropic disclosed Thursday that Claude, its flagship AI model, now leads more than a quarter of the company's research and development efforts—a milestone that underscores both the rapid advancement of AI capabilities and growing concerns about models building their own successors.
Claude currently leads 26 percent of Anthropic's model R&D work, meaning it can complete most tasks end-to-end from high-level prompts while remaining under human supervision, according to the company's announcement. The model is not operating autonomously. Additionally, Claude participates in roughly 90 percent of the company's research and development through what Anthropic describes as "collaboration"—handling substantial portions of work under close human direction.
The disclosure, first reported by the Associated Press, arrives amid intensifying debate over AI safety, with Anthropic CEO Dario Amodei among those advocating for greater caution in development.
Rapid progression toward self-improvement
The pace of Claude's expanding role is striking. In February, the model led zero percent of Anthropic's research and development work. Six months later, by August, that figure had climbed to 26 percent.
Anthropic emphasized that models accelerating their own development could make it "more challenging for humans to understand or control these systems." The company stopped short of stating how close it believes the industry is to achieving recursive self-improvement—the point at which a model can autonomously build its successor—but called for transparency around these metrics.
"We should do everything possible to minimize the gap between what frontier labs know and what the public knows," Anthropic stated in a blog post accompanying the announcement. The company urged other AI developers to share similar measurements regularly using public methodologies that allow for comparison across labs and over time.
Oversight infrastructure scales up
Anthropic also detailed its oversight measures, noting that approximately 30,000 agents were conducting research and engineering work as of August. The company said monitoring systems that detect agent misbehavior are essential as these systems scale. Anthropic recently committed to embedding external third-party evaluators within the company to monitor safety efforts.
Why it matters
The disclosure provides rare quantitative insight into how leading AI labs are approaching recursive self-improvement—a capability that could dramatically accelerate AI progress while making oversight more difficult. Anthropic's call for industry-wide transparency comes as tech leaders remain divided on whether to slow development, with some supporting caution while others, including President Donald Trump, have pushed back against constraints. For enterprise leaders evaluating AI partnerships, understanding how close providers are to autonomous model development has direct implications for risk management and governance strategies.
The announcement follows the resignation last week of an Anthropic researcher who issued warnings about AI threats to humanity, sparking renewed discussion among tech leaders including Amodei, OpenAI CEO Sam Altman, and Elon Musk about the pace of AI development.
Details were first reported by the Associated Press.
