Anthropic analyses 207,000 research sessions to see how scientists use Claude

Anthropic published a report on how researchers use its products, led by Johannes Hermle, Ben Jones and Lukas Althoff, based on 207,049 Claude and Claude Code sessions from consumer plans between 2 August and 30 September 2026. Of these, 128,914 came from the Chat and Cowork interfaces and 78,135 from Claude Code.

A language model labelled each session by scientific field and research task, using a fixed taxonomy; Anthropic says those labels were the only information from the conversations that the research team retained, and that results are reported only in aggregate. That means the study measures where scientists use Claude, not whether the work was good.

The headline findings are about distribution. Relative to the number of researchers in each field, use is higher in computer science, mathematics and statistics and lower in health and the lab-based physical sciences. Data analysis accounts for about a third of Claude sessions and nearly half of Claude Code sessions; idea generation and data gathering are used less, although scientists in an accompanying survey of 528 researchers (mostly life and health scientists who already use AI) ranked those as the most important tasks for discovery. On Claude Science, a specialised product, idea generation makes up about twice the share it does on Claude. Anthropic also built a measure of how well AI can in theory perform each task, and reported that the tasks best suited to AI get two to four times the use that researcher time would predict.

The report then turns to scenarios, and its own wording is cautious: they are “illustrative scenarios, not forecasts”. Wide adoption of current capabilities could, it says, double research output in some fields. If AI fully automated all non-physical research work, output in computer science and statistics could rise as much as 15-fold, against three- or fourfold in chemistry and biology, where physical bottlenecks remain. The 15-fold figure is a modelled outcome built on an economic production function and the authors’ assumptions, not a measurement. The authors list limits: they ignore constraints on shared facilities, the cost of experimental equipment and the quality of ideas.

On expectations, about 40% of the surveyed researchers said they expect AI to carry out full research projects on its own by 2030, about 64% by 2035, and 16% said never. Since the sample is self-selected AI adopters and the data come from one company’s own products, the figures describe Claude users rather than science as a whole.

The report came a day after Anthropic’s vault-particle preprint and follows its September enzyme-discovery claim; those are individual case studies, while this one measures routine use.