Anthropic reveals that Claude is already involved in more than 90 % of its artificial intelligence research and development tasks. In a quarter of these tasks, the system can take over the work—from a general instruction through to completing a large part of the process—always under human supervision.
Until very recently, the idea seemed like something out of science fiction: artificial intelligence helping to build the next generation of artificial intelligence. But this is no longer just a theoretical possibility. Anthropic, the company behind Claude, has just released internal data showing the extent to which its own models are contributing to the development of future systems. Anthropic
The numbers are striking. According to data from August 2026, Claude participates in more than 90 % of Anthropic's AI research and development work at a level that the company classifies as "collaboration" or higher. Even more significant: in the 26 % of those tasks—Claude is already “leading” the work. Anthropic
What exactly does “lead” mean?
This does not mean that Claude has decided on its own to create a new artificial intelligence, nor that Anthropic has handed over control of its labs to a machine. The company itself defines this level as one in which AI can receive a high-level instruction and carry out most of a task from start to finish, while a human continues to oversee the process. Anthropic also states that Claude does not operate completely autonomously in any of the research and development areas it assessed. Anthropic
The difference is significant.
What is extraordinary is not that an AI has become independent, but the speed with which it is evolving from a tool used by human researchers to a direct participant in the process by which future models are developed.
FROM ASSISTANT TO COLLABORATOR
For years, artificial intelligence tools were mainly used to help with specific tasks: completing code, finding errors, summarizing documents, or suggesting solutions. That landscape is changing.
Anthropic says that in May of this year More than 80 % of the code incorporated into his production code could be attributed to Claude. The company also notes that its engineers' productivity—measured in lines of code added daily—has increased significantly since the models began performing tasks over longer periods rather than simply suggesting code snippets. Anthropic
Now this phenomenon is affecting research itself.
To measure this, Anthropic developed an internal index that classifies tasks from level zero—with no AI intervention—to level five, where a system would operate autonomously without human involvement. Claude currently operates at different levels depending on the task, but the company considers the growth in tasks where the model can manage a large part of the process under supervision to be particularly significant. Anthropic
The trend has been rapid: the share of research and development that Claude “led” was less than 1 % in February and reached 26 % in August. Anthropic
30,000 AGENTS AT WORK
There is another piece of information that helps put this into perspective.
Anthropic reports that, in August, approximately 30,000 AI agents were simultaneously conducting research and engineering work within its most widely used internal platform. These are not 30,000 independent intelligences, but rather multiple instances of agents based on its systems working on different tasks. Anthropic
The company states that these agents are subject to monitoring mechanisms. Some monitors operate in real time and can block certain actions; others analyze activity after the fact to detect problematic behavior or possible signs of misalignment. Anthropic
That is precisely where the issue lies that makes this news story much more than just a simple productivity gain.
CAN AN AI IMPROVE ITSELF?
There is a concept known as Recursive Improvement in Artificial Intelligence: A sufficiently advanced system helps create another, better system; that new system then contributes to the creation of an even more capable one, progressively accelerating the cycle.
Anthropic believes that we haven't reached that point yet.
In fact, the company acknowledges significant limitations in its own measurements. Among these is the fact that it uses Claude models to evaluate some of the work produced by other Claude models, which can introduce shared errors. For this reason, it suggests that these types of metrics could be verified in the future by independent third parties. Anthropic
Nor is there any evidence in these data that an AI has taken control of its own development. Humans continue to set goals, provide infrastructure, establish limits, and monitor the results.
But the border is shifting.
In another experiment reported by Anthropic, agents powered by Claude worked on an open-ended AI safety problem, proposing hypotheses, conducting tests, exchanging results, and iterating on them. Humans had still chosen the problem and defined how to evaluate the results, a fundamental limitation on interpreting the experiment as full autonomy. Anthropic
A QUESTION THAT NO LONGER BELONGS TO THE FUTURE
What is truly significant about this announcement may not be the 26 %, or even the 90 %. Rather, it is that one of the leading artificial intelligence companies considers it necessary to begin publicly measuring To what extent does AI contribute to the creation of new AI?.
The question is no longer simply about what these models will be capable of in five or ten years.
Now there is another, much more pressing question: What happens when an increasing portion of the progress in artificial intelligence begins to be driven by artificial intelligence itself?
Anthropic argues that making these metrics public will make it possible to track the pace of this process and detect when systems are approaching much higher levels of autonomy. It also proposes common methodologies and external evaluations that would allow for comparisons of what is happening across different laboratories. Anthropic
For now, Claude needs humans. He needs goals, supervision, infrastructure, and boundaries. He is not building his successor on his own.
But something has changed.
Artificial intelligence is no longer just the product that comes out of the lab. It is also becoming one of the tools used within the lab to build what comes next.
