Talk about AI in science tends to dwell on breakthroughs: protein structures, new crystals, solved maths problems. Anthropic's new study of how researchers actually use Claude paints a more ordinary picture. Scientists mostly use it to analyse data and to write, and they use it least for the tasks they themselves say matter most for discovery.
The study covers 207,049 research sessions on consumer plans between 2 August and 30 September 2026: 128,914 in Claude's chat and Cowork interfaces and 78,135 in Claude Code. A language model classified each session by field and task against a taxonomy of 7,262 research tasks across 34 fields, keeping only those labels; Anthropic says no research content was extracted and results are reported only in aggregate. Data analysis alone made up about a third of sessions in Claude and nearly half in Claude Code. Computer science led on Claude with 11.9% of sessions, then engineering at 8.4% and medicine at 7.7%. On Claude Code the skew is sharper: computer science accounts for 4.3% of the world's researchers but 32.7% of sessions.
The more interesting finding is the match between use and ability. Anthropic built a measure of how well AI could, in principle, perform each research task. The top fifth of tasks on that measure got two to four times as much use as the time researchers spend on them would predict, while the bottom fifth, mostly physical lab work, got very little. Idea generation and data gathering, which surveyed scientists rank as most important, were used less; on Claude Science, Anthropic's specialised research product, idea generation's share is about twice that on Claude.
For readers of our Claude and Claude Code reviews, this fits what those products are good at: working through data, code and drafts quickly. It also suggests the biggest gains are going to the most computational fields first, including the ones that build AI.
The forecasts deserve caution. A simple model in the report suggests current capabilities, widely adopted, could double research output in some fields, and that automating all non-physical research work could raise output in computer science and statistics as much as 15-fold, with bench sciences limited to three or four times by physical bottlenecks. Those are scenarios, not measurements. About 40% of surveyed active AI users expect models to run full research projects alone by 2030, which is opinion. And the whole study is Anthropic analysing its own product's consumer users, which leaves out enterprise and API work.