207,049 Claude Research Sessions Show Analysis Ahead of Ideas

Anthropic published a study of 207,049 scientific research sessions on November 5, finding that scientists in its sample used Claude more for analyzing data and writing than for generating hypotheses or collecting information. The new company research report offers a detailed view of work performed on one provider’s consumer products—not a census of scientists or proof that AI has already accelerated discovery by a particular amount.

The main sample covers August 2 through September 30, 2026: 128,914 sessions from Chat and Cowork, grouped as “Claude,” and 78,135 from Claude Code. A session is an interaction, not a unique person, a finished paper or a verified scientific result. Keeping those units separate is essential when interpreting the headline total.

By the numbers: analysis dominates the recorded work

Data analysis accounted for 33.7% of Claude research sessions and 48.1% of Claude Code research sessions. Ideation and hypothesis generation accounted for 4.2% and 2.1%, respectively. TBR calculations show a 14.4-percentage-point larger data-analysis share on Claude Code, while the ideation share on Claude was twice that on Claude Code. These are differences in each product’s usage mix, not evidence that one product makes a scientist twice as creative.

The chart shows only those selected task categories; its bars are not meant to sum to 100%. Different users and tasks can select into different products. The report also discusses estimated researcher time from a separate task framework, but those estimates are not merged into this chart as if they were observed session counts. Neither the time model nor the usage shares measure the quality of a resulting scientific finding.

How the company classified scientific conversations

Anthropic says language-model classifiers first identified scientific work and then assigned sessions to fields, subfields and tasks. The SciNet research taxonomy contains 7,262 tasks across 34 fields and 318 subfields. Some activities, such as manuscript drafting, apply across many disciplines; others are highly specialized. This structure makes different kinds of work visible without treating every use of code as computer-science research.

According to the report, the research team retained fixed category labels rather than open-ended details of hypotheses, datasets or unpublished findings. Results were aggregated and subject to minimum-account thresholds. The accompanying methodological appendix provides sample-construction and classifier details. Those privacy measures do not turn model-generated labels into perfect observations: classification error and selection into Claude use remain limits on what can be inferred.

A survey adds perspective, not a representative global count

The researchers also surveyed 528 scientists, primarily in life and health sciences, who were active AI adopters. Respondents were recruited from applicants to Anthropic’s science-related grant programs and affiliated with institutions in the United States and Canada. That recruitment path matters: people applying for AI research credits are likely to use these systems more intensively than the typical researcher.

Respondents ranked idea generation and data gathering as especially important for discovery even though those activities were less prominent in the observed AI use. Some data gathering requires physical experiments or patient interactions. Hypothesis generation is harder to classify and can be embedded within other tasks. The report therefore does not establish that low session share means scientists consider an activity unimportant or that AI is incapable of contributing to it.

The acceleration estimates are scenarios, not results

Anthropic separately models how research output might change if AI capabilities were widely adopted or more non-physical work became automated. Its assumptions include the costs of researcher and AI time and the way different tasks complement one another. These illustrative scenarios are not observed increases in publications, successful experiments or medical discoveries. Physical laboratories and scarce facilities can remain bottlenecks even when text and code work gets faster.

The provider has a commercial interest in demonstrating valuable uses, alongside the methodological contribution of releasing aggregate evidence. TBR’s reporting on Anthropic’s financial expansion supplies business context, while the AI research section follows scientific claims and evidence. The study’s treatment of conversation data also sits alongside the separate debate over industry-wide AI training opt-outs; publication-data rules and private research-session analysis are not the same consent problem.

For research institutions, the useful finding is narrower than a promise of autonomous science. The evidence identifies where this sample already turns to AI and where important work remains outside it. Whether that pattern produces better discoveries requires outcomes and validation that session counts alone cannot supply.

Sources

1. Anthropic scientific-use report, November 5, 2026
2. Anthropic methodological appendix
3. SciNet public research-task taxonomy, checked November 5

Additional sources

• Wikimedia Commons photograph record and licensing information