Anthropic’s report on Claude in scientific research found that data analysis and writing made up the largest shares of classified Claude and Claude Code research sessions. The report, published November 5, 2026, analyzes sessions collected from August 2 through September 30, 2026. Its larger research-acceleration figures come from hypothetical model scenarios, not observed changes in scientific output. Anthropic’s report covers usage patterns, a researcher survey, and those modeled scenarios.
The session data describe where Claude was used; the acceleration figures describe what a model estimates under specified assumptions.
What Anthropic studied
The main analysis covers 207,049 consumer-plan sessions classified as scientific research: 128,914 Claude Chat and Cowork sessions and 78,135 Claude Code sessions. Anthropic used classifiers to assign research fields and tasks according to SciNet, a taxonomy spanning 7,262 tasks, 34 fields, and 318 subfields.
Anthropic says its classifiers retained category labels rather than open-ended research content, and that results were aggregated subject to minimum-account thresholds. The figures therefore describe classified product sessions, not a record of scientific findings made with Claude.
A separate survey covered 528 researchers, primarily in life and health sciences at institutions in the United States and Canada. Anthropic recruited respondents mainly from applicants to its AI-for-science and related grant programs. The company says those applicants are likely to use AI more intensively than typical scientists, so the survey does not represent researchers as a whole. Respondents ranked ideation and hypothesis generation as most important for producing significant research results, with data gathering second, even though both categories accounted for smaller shares of sessions.
Data analysis leads Claude and Claude Code sessions
Anthropic classified data analysis as 33.7% of Claude sessions and 48.1% of Claude Code sessions. Writing and communication accounted for 32.0% and 17.6%, respectively. The estimated shares of researcher time are a separate benchmark, not a measure of how much time those sessions took.
| Research task | Claude sessions | Claude Code sessions | Estimated researcher-time share |
| Data analysis | 33.7% | 48.1% | 27.6% |
| Writing and communication | 32.0% | 17.6% | 16.0% |
| Data gathering | 11.5% | 11.9% | 28.6% |
| Ideation and hypothesis generation | 4.2% | 2.1% | 7.1% |
The difference is clearest for data gathering: it represented more than a quarter of Anthropic’s estimated researcher time, but about 12% of sessions on either product. Ideation also took a larger share of estimated researcher time than of sessions. In the survey, researchers rated these activities as important, while Anthropic notes that physical work can constrain data gathering and that ideation can be difficult for classifiers to categorize.
Claude Science sessions focus on life sciences
In Anthropic’s separate Claude Science analysis, biomedical sciences made up 31.1% of research sessions, biology 24.1%, and medicine and clinical sciences 8.7%.
For sessions in biology and biomedical sciences, Anthropic estimated human-only task time at 4.9 hours for Claude, 5.9 hours for Claude Code, and 10.3 hours for Claude Science. It classified doctoral-level content in 48.0%, 67.7%, and 82.7% of sessions, respectively; its high-autonomy measure was 23.7%, 64.1%, and 73.3%. These are characteristics assigned to sessions, and Anthropic says differences in who adopts Claude Science may partly account for the pattern.
Acceleration figures are scenario outputs
Anthropic’s first scenario assumes broad adoption of current AI capabilities. Under that assumption, its model estimates that AI could take on 27% to 52% of researcher time across fields, with a median of 38%. Modeled output acceleration ranges from 1.35× in arts to 2.02× in statistics; the report gives 1.41× for chemistry and 1.96× for computer science.
Two more expansive scenarios assume AI automates 75% or 100% of non-physical research steps, while researchers continue to perform physical steps. In the 75% scenario, the model estimates about 8× acceleration in computer science, 7× in statistics and library and information science, and 6× in economics. In the 100% scenario, it estimates roughly 15–16× in many fields with few physical tasks, compared with about 3–4× in bench sciences such as biology, chemistry, and materials science.
These figures are conditional outputs of an illustrative model. Anthropic’s analysis excludes constraints including facility capacity, machinery and experimental-input costs, and the quality of research ideas. Physical work remains part of the modeled process, which is why the scenarios yield smaller estimates for bench sciences.
Keep literature claims tied to their papers
A practical way to organize a literature review is to keep each topic’s papers, reference files, and instructions together. In a tutorial, Andy Stapleton demonstrates using separate Claude Projects for research materials and custom instructions.
A commenter in a research-workflow discussion recommends asking Claude for the exact passage supporting each claim, then checking that passage against the paper. Other users warn about findings attributed to the wrong study and about mistaking an unsuccessful search for a genuine gap in the literature. These are individual reports, not measured error rates. Researchers’ discussion of source-checking workflows points to a concrete habit: match each summary claim to the source text before carrying it into a review.
