Anthropic's economics team published a report on November 5 that tallies what researchers actually do with Claude. The sample covers 207,049 sessions on consumer plans that were tagged as research-related between August 2 and September 30. Of these, 128,914 came from Claude (including Chat and Cowork) and 78,135 from Claude Code, with an additional supplementary sample from Claude Science.
Sessions were classified with SciNet, a taxonomy the team built in-house, covering 7,262 tasks, 34 disciplines and 318 subfields. According to the report, the classifier outputs only discipline and task labels and does not extract research content; results are published in aggregate only, and cells with too few accounts are suppressed.
Most use goes to crunching data and writing
By share of sessions, data analysis accounts for 33.7% of Claude use and 48.1% of Claude Code use. Writing and communication makes up 32% of Claude sessions and 17.6% of Claude Code sessions. The single most common task is drafting a paper manuscript, at 18.6% of Claude sessions.
Set against how researchers actually spend their time, the gap is clear:
| Task | Claude | Claude Code | Share of research time |
|---|---|---|---|
| Data analysis | 33.7% | 48.1% | 27.6% |
| Writing and communication | 32.0% | 17.6% | 16.0% |
| Data collection | 11.5% | 11.9% | 28.6% |
| Ideation and hypotheses | 4.2% | 2.1% | 7.1% |
Data collection takes up nearly 30% of researchers' time but only a little over 10% of sessions, most likely because sampling and running experiments require hands-on work. Ideation and hypothesis generation was rated the most important stage by the scientists surveyed (35.6 out of 100), yet it sees the lowest usage.
The spread across disciplines is also lopsided. Computer science researchers make up just 4.3% of researchers worldwide but 32.7% of Claude Code sessions; researchers in AI and machine learning subfields are 1.2% of the global total but 15.2% of Claude Code use. Experimental fields such as chemistry and medicine show lower usage. As the report puts it:
"In short, scientists appear to use AI where it works best."
Working out the speedup
The report offers estimates under three scenarios. The first assumes only broad rollout of current capabilities: AI could take on 27% to 52% of research time, with a median of 38%, speeding output by roughly 1.35x (arts) to 2.02x (statistics) across disciplines, about 1.96x for computer science and about 1.41x for chemistry. The third scenario assumes AI handles every non-physical step, which would let statistics, computer science, economics and mathematics reach 15x to 16x, while biology, chemistry and materials would gain only 3x to 4x, because physical steps make up 38% to 41% of research time in those fields.
These multiples rest on a cost assumption: AI at $30 per hour, researchers at $95 per hour, and 0.2 hours of AI work for every researcher hour saved. By that rough math, each hour of human labor saved costs only about $6 in AI spend, so cost is barely a constraint. The bottleneck is the hands-on work in the lab. The report itself states, "Physical bottlenecks remain powerful."
The authors stress that the second and third scenarios are hypotheticals, not forecasts, and that the estimates leave out equipment, reagents and lab capacity. Both the classification and the time estimates were produced by a large language model, so figures at the subfield level are noisier.
What the scientists themselves think
An accompanying survey gathered responses from 528 scientists, mostly in the life and health sciences and concentrated in the United States and Canada. They were applicants to Anthropic's science program, so the sample leans toward enthusiastic users. 38% of respondents believe AI can already match top researchers at data analysis, and they expect that share to reach 94% by 2030. About 40% think AI will be able to complete an entire research project on its own by 2030, while 16% say it never will.
The Claude Science user profile is more specialized: biomedicine and biology together make up more than half, 82.7% of sessions reach PhD-level content, and 73.3% involve highly autonomous execution. These users are more expert to begin with, so the differences cannot all be credited to the product, a point the report makes as well.
Sources: Anthropic research report “How researchers use Claude”, CocoLoop, with session samples, task shares, discipline breakdowns, the three-scenario speedup multiples and cost assumptions checked against the report text; survey figures are as disclosed in the report.