Anthropic put a number on a question the AI industry has been arguing about in the abstract. In a report released Thursday, the company said that its Claude chatbot now “leads” 26 percent of its artificial-intelligence research and development work, up from nearly zero at the start of the year. The company also said Claude assists in roughly 90 percent of its employees’ work.
The disclosure is an attempt to make the progress of AI measurable in public. As concern has grown that AI systems could rapidly improve their own capabilities and move beyond human control, Anthropic has argued that the public needs a way to track how much of the work of building AI is now being done by AI itself. The report is the company’s latest contribution to that debate.
The 26 percent figure is the headline, but the scale behind it is what gives it meaning. Anthropic divides the degree of autonomy in a task into five levels, ranging from work done entirely by humans to work handled entirely by AI with people only watching. “Leading” a research task means the model is doing the substantive work of a project, not merely assisting around the edges.
The company also said it plans to bring in third-party evaluators from multiple institutions and grant them the same access to internal processes, systems, and data that its own employees hold. The move is a bid to give outside researchers a way to verify the company’s claims rather than take them on faith.
The language in the report is direct about the stakes. “AI systems are becoming more powerful at an exponential rate and have begun to automate more of the process of building themselves,” the company wrote, adding that as the world considers slowing the development of frontier AI, the public needs more information. The framing places Anthropic’s own disclosures inside a debate it has been actively shaping.
The company has been one of the loudest voices calling for caution about the pace of AI development, and the report reads as evidence for its argument as much as a transparency measure. A model that leads a quarter of the research effort of the company that built it is, by Anthropic’s own framing, a step toward the self-improving systems that the cautionary voices describe.
The outside reaction was quick to sharpen the point. Agence France-Presse ran the day’s news under a blunt headline: AI systems are heading toward building themselves. It is a compression of the report’s implications, and one that Anthropic did not dispute.
The metrics also serve a commercial purpose. Anthropic is competing for the trust of enterprises that are deciding whether to hand important work to AI, and a published account of how much of its own work Claude now performs doubles as a demonstration of capability. The company that runs on its own product is offering itself as proof of what the product can do.
The five-level scale is the intellectual scaffolding of the report. At the bottom sits work done entirely by people; at the top sits work handled entirely by the model, with humans present only to watch. “Leading” a task occupies the upper band of that scale, meaning the model is doing the substantive work while a person directs and reviews. It is a definition chosen to be measurable, and one that invites argument about where the lines are drawn.
The disclosure is unusual in its specificity. Most AI developers describe their models’ capabilities in benchmarks and demos, not in the share of their own research a model now performs. Anthropic’s choice to publish the number is a bet that transparency will be rewarded, both by regulators and by the enterprises deciding whether to trust the technology with their own work.
Skeptics have already noted that self-reported metrics are only as good as the definitions behind them, and that a company with a position in the safety debate has an interest in publishing numbers that support it. The third-party evaluators Anthropic has promised are the mechanism for addressing that skepticism, but until they are named and given access, the 26 percent figure stands on the company’s own account.
The choice to publish now is also a response to the moment. Governments are weighing whether to slow the development of frontier AI, and the companies are arguing about whether such systems can be controlled. A report that quantifies how much of its own construction AI has begun to handle is a contribution to that argument, and a way for Anthropic to put evidence behind a position it has taken publicly for years.
The challenge for readers of the report is that the categories are still young. A five-level autonomy scale with no agreed definitions leaves room for disagreement about what “leading” a task really means, and the third-party evaluators have not yet been named or given access. What Anthropic has published is a starting point, a set of numbers that will mean more once independent researchers can test them against their own observations.
