Scientists Use Claude Mostly for Data Analysis and Writing, Anthropic Study Finds

Anthropic economists and researchers studied 207,049 scientific sessions on consumer plans between August 2 and September 30, 2026. Of these, 128,914 were on Claude and 78,135 on Claude Code. The company said it used privacy-preserving methods: a classifier tagged each session by field and research task, only those tags were analysed, and results were reported only where enough distinct accounts were involved.

The report gives a rare look at where AI is already doing research work and where it is not. That matters for anyone deciding how to fund science or deploy AI in R&D. Anthropic said it also wanted to estimate how far AI could speed up scientific discovery.

Use Tracks What the Models Do Well

Data analysis made up 33.7% of Claude sessions and 48.1% of Claude Code sessions, compared with 27.6% of the time researchers spend on it. Writing and communication took 32.0% of Claude sessions, twice its 16.0% share of researcher time. Data gathering, which fills 28.6% of researchers' time, accounted for only about 11.5% to 11.9% of sessions. Ideation and hypothesis generation made up 4.2% of Claude sessions and 2.1% of Claude Code sessions. That was below its 7.1% share of researcher time, even though the scientists surveyed ranked it as the most important category for producing important results.

Relative to how many researchers work in each field, use was heaviest in computer science, mathematics and statistics, along with the social sciences and economics. Computer science accounts for 4.3% of the world's researchers but 32.7% of Claude Code sessions. Lab-based physical sciences such as chemistry, and health fields, used Claude less. The authors built a measure of how well AI can in theory perform each research task. They found that the tasks scoring in the top fifth received two to four times as much Claude use as researchers' time on them would predict. "In short, scientists appear to use AI where it works best," they wrote.

Anthropic's research-focused Claude Science product, which connects the model to scientific tools and databases, showed a different pattern. Life sciences made up over half of its sessions. Ideation took about 8% of use, roughly in line with how much time researchers spend on it. In biology and biomedical sessions, 73.3% showed high AI autonomy, against 23.7% on standard Claude.

From Double to Fifteenfold

Using a simple model, the authors estimated that if current capabilities were widely adopted, AI could take on between 27% and 52% of research time. Output would rise from 1.35 times in the arts to 2.02 times in statistics, with computer science at 1.96 times. If AI could do all non-physical research work, output in computer science and statistics could rise about 15-fold. In bench sciences such as biology and chemistry, physical lab work would hold the gain to three to four times.

The report calls these scenarios illustrative rather than forecasts. They assume AI agents cost $30 an hour against $95 for a researcher, and they leave out equipment, facilities and the quality of ideas. The authors also note that the fields with the most room to speed up include those that build AI itself.

The accompanying survey covered 528 scientists, mostly in the life and health sciences, recruited from applicants to Anthropic's AI for Science grant programme at US and Canadian institutions. Anthropic acknowledged they are probably heavier AI users than most researchers. About 40% expect AI models to carry out full research projects on their own by 2030, about 64% by 2035, and 16% said this would never happen.