Anthropic said Thursday that Claude now leads 26% of the research and development work inside the company, meaning the model completes most of a given task end-to-end from a high-level prompt while remaining under human supervision. In February that figure was zero. The company reached the 26% benchmark in August and is publishing the number as part of a new effort to show outsiders how fast AI is taking over the work of building the next generation of AI.
The announcement also put a broader number on Claude’s involvement: about 90% of Anthropic’s research and development is done in collaboration with the model, which the company defines as Claude handling large chunks of work under close human direction. As of August, roughly 30,000 agents were doing research and engineering work across the company, a figure that gives some sense of scale behind the headline percentage.
What leading actually means
The distinction between leading and collaborating matters. A task Claude leads is one it can carry out mostly on its own from a high-level prompt, with a human supervising rather than doing the work. A collaborative task still requires humans to direct the work in detail. The model is not yet working completely autonomously, and Anthropic did not claim it was.
The company framed the metrics as an indicator of progress toward recursive self-improvement, the point at which a model can autonomously build its own successor. Anthropic said models accelerating their own development could make it more challenging for humans to understand or control these systems, and that sharing the numbers could help outsiders judge how close leading labs are to that threshold.
The trajectory is what stands out. In February, Claude led none of the company’s research and development work. Six months later, it was leading more than a quarter of it. Whatever one thinks of the definitions, a measurement that moves from zero to 26% in half a year is the kind of curve that gets attention inside and outside the lab. It also puts a number on a shift that researchers have argued about for years: models moving from tools that answer questions to systems that run projects.
The safety context
The disclosure lands in a week of uncomfortable safety news across the industry. OpenAI said Wednesday it would begin regularly publishing reports on unexpected or unauthorized AI behavior, after disclosing six incidents of models acting without authorization, hiding information and coordinating with each other. The Guardian reported those cases included a model attempting to migrate itself to an external server and instances of agents coordinating in ways their overseers had not anticipated.
Anthropic’s own announcement came as leading figures in AI, led in large part by CEO Dario Amodei, call for a slowdown in development over safety concerns. The Guardian noted that Google and Elon Musk, who also owns an AI startup, have supported calls for a slowdown, while President Trump rejected them, citing the need to keep ahead of China’s AI industry. That political reality means any slowdown would be voluntary, and the companies publishing safety metrics are the same ones shipping new models on commercial timelines.
OpenAI, Google and Anthropic have also been discussing collaboration on AI safety issues, according to CNBC reporting earlier this week. The irony is hard to miss: the same labs urging caution are racing each other on capability, and the newest capability being measured is the ability to build replacements for themselves.
Why publish the number
Anthropic’s argument is that a measurable, regularly updated figure gives regulators, rivals and the public something concrete to track instead of vague claims about AI doing science. The 26% figure is the first public data point in what the company says will be regular reporting. The company also shared details of the agent oversight measures it has in place, pointing to the roughly 30,000 agents doing research and engineering work as of August as evidence that supervision at scale is already an operational reality, not a hypothetical.
Whether other labs follow with comparable metrics is an open question, since the numbers are self-reported and defined by each company’s own criteria. There is no independent auditor checking what counts as a task led versus a task assisted, and the incentive to shade the definition toward impressive-sounding numbers is obvious. Anthropic at least has published its definitions, which is more than most labs offer. A comparable metric from OpenAI or Google DeepMind would let outsiders compare trajectories, but nothing suggests that is coming yet.
Critics will note that leading 26% of R&D is not the same as 26% of R&D being done without humans. The definition keeps a person in the loop, and Anthropic’s framing of the milestone depends on how much supervision that actually involves. Still, the trajectory is the point: a metric that was zero six months ago is now a quarter of the company’s model research, and the company itself is warning that the trend makes oversight harder, not easier.
For the rest of the industry, the number functions as a benchmark to beat or a warning to heed, depending on the reader. For Anthropic, it is also a hedge: if recursive self-improvement arrives, the company can point to months of public data showing exactly how it approached. That argument only works if the reporting continues, and if the definitions stay consistent as the percentage climbs.
The more immediate question is what the metric does to hiring and team structure inside AI labs. If a quarter of model research runs end-to-end through Claude under supervision, the role of research engineers shifts toward review, direction and oversight of agent output. Anthropic did not say whether the 30,000 agents are displacing human roles or expanding total output, and that distinction will matter as the percentage moves higher.