Scientists use Claude disproportionately for data analysis and writing, while data gathering and idea generation draw smaller shares of activity, according to a technical report Anthropic published on November 5, 2026. The findings describe research workflows on the company’s products, without directly measuring improvements in scientific productivity or discovery quality.
The observational analysis covered 207,049 scientific sessions on consumer plans from August 2 to September 30, 2026: 128,914 Chat and Cowork sessions, grouped as Claude, and 78,135 Claude Code sessions. Language-model classifiers assigned fields and tasks using SciNet, which covers 7,262 tasks across 34 fields and 318 subfields. Anthropic says the research team retained predefined category labels rather than research-content details, and reported only aggregates meeting minimum distinct-account thresholds.
Data analysis represented 33.7% of Claude sessions and 48.1% of Claude Code sessions, compared with an estimated 27.6% of researcher time. Writing and communication accounted for 32.0% of Claude sessions, against 16.0% of estimated time. Those time benchmarks were derived using language models, rather than direct time tracking of the sampled users.
Data gathering accounted for 11.5% of Claude sessions and 11.9% of Claude Code sessions, versus 28.6% of estimated researcher time. Ideation represented 4.2% and 2.1%, respectively, versus 7.1% of time. Physical data collection was particularly underrepresented. Across fields, computer science accounted for 32.7% of Claude Code scientific sessions, compared with an estimated 4.3% of researchers in the comparison population.
To examine how usage relates to capability, the authors asked a language model to compare pairs of research steps and combined its judgments using a Bradley–Terry statistical model. Tasks in the highest-scoring fifth received about twice the session share expected from their estimated time requirements on Claude and four times that share on Claude Code. These associations concern theoretical capability scores, not independently measured task-success rates, and do not establish that capability caused adoption.
A survey of 528 scientists broadly matched the usage pattern. Respondents ranked ideation and data gathering as the most important activities for significant findings, despite using AI less for them. Participants were primarily active AI adopters in life and health sciences, limiting how widely their responses can be generalized. About 40% expected AI to conduct full research projects independently by 2030; this reflects expectations rather than demonstrated capability.
In a supplemental Claude Science sample, ideation accounted for 8% of sessions. Specialized interfaces may therefore support different workflows, but product comparisons were not randomized and could reflect differences among users. The supplemental sample size was not supplied.
The report also explores illustrative acceleration scenarios. Under its assumptions, widespread adoption of current capabilities could approximately double research output in some fields. If all nonphysical research work were automated, output could rise up to 15 times in some formal sciences, compared with three to four times in laboratory-heavy sciences where physical work remains a bottleneck. These are model-dependent scenarios, not observed gains or forecasts. They include AI costs and do not model contributions from better research ideas.
Sessions do not measure how many scientists have adopted AI or how much science they produce. Repeated sessions within accounts may be dependent, and independent researcher counts were not supplied. Classification, time estimates and capability rankings rely substantially on language-model judgments; shared measurement assumptions could contribute to the associations. The unavailable appendix limits assessment of validation, uncertainty and model sensitivity. Peer review and independent replication of this company-reported research are not established by the supplied material.
Source: Anthropic, How scientists use Claude, November 5, 2026. https://anthropic.com/research/how-researchers-use-claude