How Researchers Use Claude

Anthropic studied 207,049 research sessions on Claude and Claude Code and found that scientists use the AI mostly where it is rated most capable, and that this is not where they think discovery happens. The report covers consumer-plan sessions from August 2 to September 30, 2026. A language model labeled each session by scientific field and research task, and the researchers kept only those labels. Data analysis accounts for 33.7% of Claude sessions and 48.1% of Claude Code sessions, against 27.6% of researchers' time. Idea and hypothesis generation gets 4.2% and 2.1%, against 7.1% of researcher time, and data gathering is also underused. In a survey of 528 scientists, those two neglected categories ranked first and second in importance for producing important results.

To test whether use follows ability, the team built a "theoretical AI capability" score. A language model compared pairs of individual research steps and judged which one AI performs better, and the comparisons were combined into a ranking, much as chess ratings come from head-to-head results. Tasks in the top fifth of that score draw two to four times as much Claude use as the time researchers spend on them would predict. Tasks in the bottom fifth, including most physical lab work, are barely used. The score is a model's judgment of what AI can do, not a measured test, but it was built without usage data and still predicts usage across tasks and fields.

That pattern bounds the near-term payoff. In Anthropic's illustrative model, wide adoption of current capabilities could take on 27% to 52% of the researcher time a study needs, raising output by about 1.35x in the arts up to about 2x in computer science and statistics. In chemistry, materials science and similar lab fields, 38% to 41% of researcher time goes to physical steps that AI cannot take on, which caps their gains even in a hypothetical where AI automates all non-physical work. The authors call these scenarios illustrative, not forecasts. The usage data covers consumer plans only, and the surveyed scientists were applicants for AI research credits, likely heavier users than most.