Artificial intelligence is no longer only the product being built inside frontier labs. It is increasingly becoming part of the workforce building what comes next.
Anthropic disclosed this week that its Claude system now leads 26% of the company’s AI research and development work — meaning it can complete most of those tasks end-to-end from a high-level prompt while a human supervises. More than 90% of Anthropic’s model R&D now involves Claude at a collaborative level or higher.
The numbers are striking because they point to a shift that AI researchers have discussed for years: advanced models helping engineers create more capable models, potentially compressing the time between generations of AI. But the details matter. Claude is not autonomously designing and releasing its own successor, and Anthropic says none of the measured R&D work has reached full autonomy.
From assistant to research partner
According to Associated Press reporting published September 18, Claude’s role has expanded rapidly during 2026. Anthropic says the share of its R&D that Claude “leads” was effectively zero in February and reached 26% by August.
Reuters reports that Anthropic had roughly 30,000 AI agents operating at any one time on its most-used internal research platform in August. The company uses monitoring systems designed to screen agent actions and intervene when necessary.
That does not mean 30,000 independent digital scientists are freely making decisions. Anthropic describes a spectrum: some work is assisted by AI, some is collaborative, and some can be led by Claude while humans remain responsible for supervision.
Why the 26% figure matters
The central issue is not whether an AI can write code or summarize a research paper. Those capabilities are already familiar. The bigger question is whether AI can increasingly perform the research, engineering, evaluation and infrastructure work required to make the next generation of AI more capable.
If that process becomes substantially automated, development could accelerate because the technology being improved would also help perform the improvement work. Researchers call the extreme version of this idea recursive self-improvement: a system capable of autonomously building a more capable successor.
Anthropic explicitly says its current measurements do not show that Claude has reached that point. Instead, the company argues that tracking AI-led R&D could help governments and the public understand how close frontier labs are getting.
The uncomfortable feedback loop
There is an obvious commercial attraction. An AI lab that automates more of its research may be able to experiment faster, fix problems sooner and reduce the human time required for complex engineering. That could improve products and potentially accelerate safety research as well.
But the same feedback loop creates a governance problem. If AI begins materially accelerating AI development, the pace of progress may become harder for regulators, outside researchers and even organizations themselves to evaluate in real time.
Anthropic acknowledges that models accelerating their own development could make systems more difficult for humans to understand or control. It is therefore proposing public measurements covering three areas: how much AI performs AI R&D, how effectively agents are monitored, and how computing resources are allocated.
Oversight is becoming part of the story
The disclosure arrives during an unusually intense debate over frontier-AI safety. OpenAI separately disclosed several cases of unexpected or concerning model behavior this week and introduced a framework for reporting such incidents. That does not establish that today’s commercial systems are escaping human control, but it shows why labs are placing more emphasis on monitoring increasingly capable agents.
Anthropic says it plans to give independent third-party evaluators access to internal processes, systems and data comparable to its internal risk teams so they can verify safety practices and monitor key development metrics.
That proposal could prove important. Self-reported numbers from AI companies are useful, but they are not the same as independent oversight. Anthropic itself notes methodological limitations, including the fact that it uses Claude in parts of the process used to evaluate Claude’s contribution.
A race with competing incentives
The controversy is sharpened by competition. Frontier labs have strong incentives to move quickly, while policymakers are being asked to consider safeguards that could slow deployment or development. At the same time, governments increasingly view advanced AI as a strategic technology with economic and national-security consequences.
This creates a difficult question: if one company slows down voluntarily, will competitors do the same? And if U.S. labs slow, could overseas rivals gain ground? Those concerns do not resolve the safety debate, but they explain why calls for coordinated standards — rather than unilateral restraint — are becoming more prominent.
What the public should take from this
The most dramatic interpretation — that AI is already independently reproducing itself — would be inaccurate. Humans still supervise Anthropic’s research process, and Claude has not reached full autonomy in the company’s measurements.
But dismissing the disclosure would also miss its significance. A frontier AI company is saying that its own model now leads more than a quarter of the work involved in developing future AI, and that AI participates in some form in more than 90% of that work.
The important trend is therefore not science-fiction autonomy today. It is the speed at which the boundary between AI being developed and AI helping perform the development is moving.
If that boundary continues to shift, the debate over AI safety may soon depend less on hypothetical future capabilities and more on measurable questions: How much development is actually automated? Who monitors the agents? How quickly can humans intervene? And who independently verifies the answers?
Sources
Anthropic — Measurements for understanding the pace of AI development inside frontier labs
Associated Press — Anthropic says Claude is helping build its next version
Reuters — Claude now leads a quarter of work building Anthropic’s next AI models
