Anthropic studied 207,000 scientific Claude sessions: 34% focused on data analysis, with ideation trailing behind

When people say AI will accelerate science, the debate usually centers on the future. Anthropic decided to examine the present: what scientists are actually doing with Claude right now. The report is based on 207,000 sessions from August and September 2026: about 129,000 from chat and Cowork, and 78,000 from Claude Code.

How it was measured

A language model reads each session and assigns labels for the scientific field, subfield, and specific task from the SciNet taxonomy. It contains 7,262 tasks, 34 fields, and 318 subfields. Conversation content is not retained: no data, hypotheses, methods, or unpublished papers. Only the labels remain, and results are shown solely in aggregate for groups with a sufficient number of distinct accounts.

One important caveat: the sample includes only consumer plans. Enterprise and academic deployments were excluded, so the picture is incomplete.

Where and why Claude is used

Computer science accounts for the largest share of sessions (11.9%), followed by engineering (8.4%) and medicine (7.7%). But the picture changes when usage is compared with the number of researchers in each field. Formal sciences—computer science, mathematics, and statistics—use Claude far more heavily than their share of the research workforce would suggest. Chemistry and medicine, where work takes place in laboratories or with patients, lag behind. The imbalance is even stronger in Claude Code: computer science represents only 4.3% of researchers worldwide but generates 32.7% of sessions.

By task, data analysis leads: 33.7% of Claude sessions and 48.1% of Claude Code sessions, although it takes up 27.6% of scientists' working time. Writing and communication account for 32% of Claude sessions versus 16% of working time. Data collection, by contrast, is underused: it consumes 28.6% of working time but represents only around 12% of sessions. This is particularly true where physical constraints intervene, such as clinical and biochemical measurements.

The paradox is that, in a survey, scientists themselves identified idea generation and data collection as the most important tasks for discovery, yet they use AI for them least often. The specialized Claude Science shifts the balance: the share devoted to ideas is roughly twice as large as in standard Claude.

Why usage looks this way

The authors estimated how well AI can, in principle, perform each task. Tasks in the top 20% receive two to four times more usage than their share of working time would imply. The bottom 20%, consisting mainly of physical work, are barely touched. AI adoption is highest where tasks are cognitive, lend themselves well to reinforcement learning, and produce verifiable outcomes.

What this means for the pace of science

Under a simple model, widespread adoption of today's capabilities could double research output in some fields. If AI takes over all nonphysical work, output in computer science and statistics could grow 15-fold, while chemistry and biology would see only a three- to fourfold increase because of physical constraints. Around 40% of surveyed active AI users expect models to be capable of running research projects entirely autonomously by 2030.

There is another effect: the fields with the greatest potential include AI development itself. This creates a feedback loop in which faster research produces more powerful models, which then accelerate research further.

Practical takeaway

For product teams and founders, the main lesson lies in the gap between what scientists do with AI and what they consider important. The open opportunity is in ideas, hypotheses, and data collection—especially where these can be handled through digital processes, such as searching archival documents and texts. For developers, the signal is clear: AI gains traction where results are easy to verify. If you are building a tool for researchers, start with data analysis and add verifiability before moving on to hypothesis generation.

The limitations also matter. Labels are assigned by a classifier model, the 15-fold increase is a hypothetical scenario rather than a forecast, and the authors themselves flag the risks: less human oversight, potentially dangerous research, and changes to scientists' careers.