The Pulse
OpenAI, Anthropic and xAI Put Recursive Self-Improvement in Sight
Anthropic says Claude now leads 26% of its model research and development while remaining under human supervision. OpenAI, xAI and Microsoft describe different paths toward AI systems that can help build their successors, raising new questi
AI.info Team ·
Anthropic says its Claude models now lead 26% of the company’s model research and development, a disclosure that has pushed recursive self-improvement from a distant theory toward an active engineering question.
The figure does not mean Claude is independently designing and training a successor. Anthropic says the work remains under human supervision, and the company has not said how close it believes it is to fully autonomous model improvement. But the disclosure shows that AI systems are already taking on larger parts of the process used to build more capable AI.
The Associated Press reported the figures and the responses from Anthropic, OpenAI, xAI and Microsoft on September 19, 2026. The AP report describes recursive self-improvement, or RSI, as a process in which an AI system helps improve itself and build the next version of the system.
Anthropic’s 26% disclosure changes the argument
Anthropic’s own account is more qualified than the most dramatic interpretations of RSI. In a paper published by The Anthropic Institute, the company says humans still choose the goals of major engineering and research projects, while Claude increasingly handles the methods needed to pursue them.
Anthropic says more than 80% of the code merged into its codebase was authored by Claude as of May 2026. The company also says an internal version of Claude improved the speed of a small model-training program by about 52 times in April 2026 when given a fixed goal and fixed correctness checks. Those results describe powerful assistance inside a bounded process, not a system that decides what kind of model should exist next.
That distinction matters. Anthropic says Claude still has a “large performance gap” when it must decide which goals are worth pursuing in the first place. The company’s reported progress therefore points toward greater automation of AI research, but not yet to a closed loop in which a model independently sets objectives, trains a successor and repeats the cycle.
OpenAI sets a March 2028 target
OpenAI says it has developed an automated “research intern” that can perform defined tasks under human direction, including work that would take a skilled researcher several days. The company says it is moving toward an automated AI researcher by March 2028.
OpenAI also warns that progress toward RSI should not be treated as an automatic good. The company says it does not yet know how to “safely get all the way to aligned, full RSI” and cannot assume that safety research will advance as quickly as model capability.
At the same time, OpenAI argues that systems capable of doing AI research could also help with alignment and safety research. Its position is therefore not a call to abandon self-improvement, but a warning that the pace and conditions of development should depend on whether people can preserve control.
xAI expects humans to leave the loop
xAI chief executive Elon Musk has offered a more aggressive timeline. According to the AP report, Musk said in March that humans were becoming less involved in improving Grok and that “every successive model is built by the one before it.” He said the process was not fully automated yet, but could reach that point by the end of 2026 and no later than 2027.
The statement describes a direction rather than a demonstrated capability. Existing systems can write code, run experiments and evaluate results, but researchers still define many of the objectives and guardrails. The unresolved question is whether a model can reliably make the high-level choices that determine which experiments deserve time, compute and deployment.
Researchers disagree about how close the threshold is
Anthony Aguirre, president and CEO of the Future of Life Institute and a physics professor at the University of California, Santa Cruz, told the AP that autonomous RSI would involve a system designing one version of itself after another. “As AI is doing more of it, it gets faster,” Aguirre said, because AI systems operate more quickly than human researchers.
John Thickstun, an assistant professor of computer science at Cornell University who studies methods for controlling AI behavior, offered a less dramatic description. He said AI has already been used for years to help create newer AI systems, including by writing code for training and infrastructure. Earlier attempts produced limited gains, he said, rather than major creative breakthroughs.
The two views are not necessarily contradictory. AI can already contribute to its own development without possessing the autonomy associated with a runaway intelligence scenario. The debate turns on how much of the research loop a model controls, whether its work generalizes beyond narrowly specified tasks and whether humans can still inspect and stop the process.
Slowdown proposals collide with competitive pressure
Anthropic says it would slow or temporarily pause frontier development if competitors at the same level did so under conditions that could be verified. The company argues that a unilateral pause would mainly hand the lead to another developer, while a coordinated pause would require several well-funded labs in multiple countries to accept the same rules.
That condition exposes the political problem surrounding RSI. Companies are being asked to slow a capability that could improve their research productivity and strengthen their position against rivals. OpenAI says the safety gap may widen as models become more capable and harder to monitor. Microsoft AI chief Mustafa Suleyman has instead described a goal of “humanist superintelligence” that remains limited and directed toward human needs.
For now, no leading lab has shown a system that independently designs, trains and deploys its own successor without meaningful human direction. The evidence shows something narrower but consequential: models are taking over more of the work used to produce future models, while the companies building them are still arguing about where supervision ends and self-improvement begins.