How scientists use Claude, and what the data cannot prove

Anthropic’s November 5 study of how scientists use Claude maps 207,049 research sessions from August 2 through September 30, 2026. Data analysis accounted for 33.7% of weighted sessions in Chat and Cowork and 48.1% in Claude Code. Writing was another major use. These are patterns of activity, not measured improvements in discovery.

For a research team deciding what to try next, the useful question is whether a particular workflow produces a result that survives scrutiny. A busy chat history alone cannot answer that question. The study is best read in three layers: usage observations, estimated task characteristics and economic scenarios.

What the sample includes

The methodological appendix limits the sample to Free, Pro and Max accounts. Claude Code sessions had to involve interactive human use. Programmatic usage was excluded. This is not a census of institutional or API research.

Language-model classifiers identify research activity, then assign each session one task and subfield. For this analysis, the research team retains category labels rather than open-ended research content, and suppresses aggregate cells with too few accounts. Readers therefore do not receive session-level evidence with which to inspect individual results.

Validation used synthetic transcripts. Field classification accuracy was 83%, while task classification reached 69%. Task accuracy fell to 64% after excluding synthetic examples that still echoed label wording. These checks assess constructed labels, not scientific correctness. The authors also report broadly stable patterns in higher-confidence subsamples.

The accompanying survey had 528 responses from 1,813 invited grant applicants, mainly in biological and medical research with a computational tilt. Applicants seeking AI credits are a selected population. Their expectations should not be presented as a representative vote of scientists worldwide.

The comparison depends on estimated research time

The underlying SciNet database, by Cristian Alamos and Lukas Althoff, organizes 7,262 tasks across 34 fields and 318 subfields. It was built mainly using language models and checked against papers, laboratory protocols and occupational data. It distinguishes hands-on researcher hours from elapsed time, which can include waiting for an incubation or a review.

That distinction matters when interpreting any research-efficiency claim. Saving ten minutes of writing does not necessarily bring an experiment’s completion ten minutes closer. A useful evaluation should track both staff effort and the date at which a trustworthy result becomes available.

What the 2x and 15x scenarios mean

Anthropic models research output at a fixed budget. Its current-capability scenario assumes widespread adoption and reaches about twice the output in some fields. A hypothetical scenario automating all non-physical research steps approaches 15 to 16 times in fields with few physical tasks. These are illustrative calculations, not observed gains or forecasts.

The model assumes $30 per AI-agent hour, $95 per researcher hour and 0.2 AI hours per researcher hour saved. Equipment costs, shared-facility constraints and the quality of new ideas are outside its calculation. Changing the assumptions changes the result.

How to use the findings without overstating them

Our practical recommendation is to choose one bounded task before treating these numbers as a planning input. A team might evaluate a data-cleaning script against a known dataset, or ask a reviewer to compare a draft methods section against the actual protocol. Specify the acceptable output before the trial.

Count review and correction time as part of the task. Keep a record of failures and reruns, and compare against the same work completed without AI assistance. If the work becomes faster but harder to reproduce, report both outcomes rather than collapsing them into a single success label.

Also separate access from evidence. ByteForward’s earlier guide to Anthropic’s science partnerships and credit routes explains support programs. Receiving credits does not validate an analysis, and participation in a usage study does not establish a tool’s suitability for a particular research protocol.

ByteForward reviewed the report, appendix and SciNet documentation. We did not inspect the underlying private sessions or independently reproduce the study. Its most useful contribution is a map of where to ask more precise questions about AI-assisted research.

Featured image: AI-generated conceptual illustration by ByteForward using the Claude mark for editorial identification. It depicts an invented workbench, not an actual laboratory or product interface. Claude is a mark of Anthropic, PBC. No affiliation or endorsement is implied.