Anthropic revealed in September 2026 that its chatbot Claude now leads over a quarter of its research and development tasks. The milestone highlights rapid progress toward artificial intelligence systems capable of recursive self-improvement, intensifying industry debates over safety oversight and the urgent need for regulatory guardrails.
Claude Leads 26 Percent of Research Tasks at Anthropic
Artificial intelligence development reached a notable milestone when Anthropic disclosed that its chatbot Claude now directs more than a quarter of the company’s internal research and engineering tasks. According to an internal review of research and development workflows, Claude successfully leads 26 percent of those projects. That figure jumped from zero in just months. As recently as February, that number was zero, the blog post said. Anthropic chief executive Dario Amodei called last weekend for leading AI companies to coordinate a slowdown in work on the technology, to allow more time to build guardrails around AI. Leaders at rival firms quickly backed the idea but President Trump criticized the proposal.
The capability indicates measurable progress toward recursive self-improvement, a scenario where artificial intelligence systems become smart enough to upgrade themselves without human intervention. While the system operates under human supervision and does not run fully autonomously, its expanding role in writing code and conducting experiments has altered how next-generation models are built. In a blog post Thursday that disclosed Claude’s contribution to Anthropic’s R&D on new AI models, the company did not say how close it judged recursive self-improvement to be. It said there was no point at which Claude operated fully autonomously
without a human in the loop today. Anthropic said that an internal review of R&D tasks found that Claude was able to “lead” that work about 26 percent of the time, with humans taking only a supervisory role and providing high-level direction.
Approximately 90 percent of Anthropic’s research and development work is conducted in tandem with Claude. The company disclosed that it has about 30,000 AI “agents,” doing research and engineering work and that it has recently directed about 6 percent of its computing power spent on R&D toward safety work. The company deployed roughly 30,000 automated agents to execute substantial portions of engineering tasks, guided closely by human researchers who provide high-level prompts. Claude’s role in research and development has grown rapidly, moving from leading 0 percent to 25 percent by August.
Computing Allocations Reveal Safety Spending Amid Growing Scrutiny
Alongside autonomous development metrics, the company detailed how it distributes its computing power. The firm reported that roughly 6 percent of the compute dedicated to AI research and development was allocated toward safety initiatives during a recent review period. The company said it used a version of Claude to measure its work on R&D using a ratings system developed by an outside AI company, before humans checked its findings. Anthropic employs several kinds of autonomous internal monitors that can escalate suspicious activity for human review.
The disclosures arrive amid intense public anxiety regarding the safety of frontier models. Last week, an employee resigned and said the company was gambling with our lives.
The anxiety has resonated across the political spectrum. Sen. Bernie Sanders (I-Vermont) appeared at an event in Washington this week where he discussed the risks of AI spiralling out of human control alongside former Trump adviser Stephen K. Bannon. Attorney General Blanche opposes AI regulation by prosecution.
OpenAI on Wednesday disclosed new examples of its AI agents cheating on tasks or going off script and released a framework under which it would disclose such incidents.
Chief Executive Dario Amodei Urges Industry-Wide Development Slowdowns
The data release accompanies a broader political and corporate debate over whether laboratories are moving too quickly. Anthropic chief executive Dario Amodei called for major artificial intelligence developers to coordinate a slowdown in building frontier systems, arguing that the industry needs more time to establish effective safety guardrails. Amodei and other prominent figures in the AI industry have recently warned about the possibility that AI will become smart enough to improve itself without input from humans. In that scenario, known as recursive self-improvement, it could become difficult for people to understand or control AI technology, Amodei and others have said.
Dario Amodei, Chief Executive of Anthropic
Amodei added that developers should release metrics publicly to help governments and independent third parties assess the pace of AI development without sacrificing commercial advantage or the United States’ lead in AI. Any frontier developer could publish these measures regularly, using a public methodology,
the company said. We are reporting these measurements because they give the public, third parties, and governments better visibility into the pace of AI development inside frontier labs.
The slowdown proposal received mixed reactions across the technology sector and Washington. While some industry leaders backed the call for caution, other executives pushed back against artificial intelligence restrictions, and political figures offered sharply divergent views on how federal oversight should proceed.
Related reading