Anthropic Counted 207,000 Science Sessions. What the Count Can’t Prove

A large count of AI-assisted research conversations can show what people ask a tool to do. It cannot, by itself, show that the tool produced a discovery. Anthropic’s analysis of 207,049 Claude and Claude Code sessions is a useful map of observed use across scientific fields, provided readers keep those two claims separate.

What was counted

Anthropic reports that its sample covers consumer-plan sessions from August 2 to September 30, 2026. Of the 207,049 sessions assigned both a scientific field and a research task, 128,914 came from Claude Chat and Cowork, and 78,135 from Claude Code. The study uses classifiers to identify scientific work and map the sessions to SciNet, a taxonomy spanning 34 fields, 318 subfields, and 7,262 tasks. Anthropic says researchers most often used Claude for data analysis and writing, with relatively more use in formal sciences and less in health and lab-based physical sciences compared with researcher counts. ^[1]

That is an unusually detailed view into a product’s observed use. It is also a view defined by the measurement pipeline. A classifier decides which sessions count as scientific, then another assigns field and task labels. A conversation can involve several tasks, and a session that looks like a research task does not establish the quality of the answer or whether it entered a published result.

The sample also has a specific boundary: consumer subscriptions and a two-month window. It does not include all institutional arrangements, private tools, work that never involved Claude, or scientists who avoid AI systems. The number of sessions is not a count of unique researchers or completed projects. ^[1]

Descriptive use is not a productivity result

Anthropic also compares inferred task duration, content complexity, and level of AI autonomy across interfaces. Those measures help characterize what kinds of work the company’s classifiers see. They are not a randomized comparison of scientists who did and did not use Claude. A session estimated to take a human several hours is not proof that the person saved that many hours. It may include planning, checking, repeated attempts, or work that would not otherwise have been done.

The distinction matters because “AI for science” can mean several different things: drafting code, summarizing literature, analyzing data, proposing a hypothesis, or helping interpret an experiment. A use count collapses those into activity. Evidence of scientific impact would need additional outcomes: independently verified findings, reproducible analyses, experiments completed, time to a defined milestone, and the cost of errors or review.

Anthropic’s report is therefore best read as an observational product-use study and a description of its classification method. It can suggest where researchers are trying the tool and help form questions for stronger evaluations. It cannot establish that AI has accelerated discovery across science. ^[1]

What researchers and institutions can do with it

For a lab deciding whether to adopt AI, the most practical lesson is to evaluate tasks separately. A literature search has different failure modes from statistical analysis or code generation. Define a benchmark from work the team already knows how to verify; record the source documents and software versions; and measure both correction time and successful completion, not just response speed.

Institutions should also ask what the study’s privacy-preserving methods mean in practice. Anthropic describes session classification rather than publishing individual transcripts, but the report’s public summary is not a full audit of data retention, selection effects, or classifier error. The company’s analysis is useful evidence about one product’s observed use, with the publisher also responsible for the analysis. ^[1]

The most durable conclusion is modest: scientists are using general-purpose and coding assistants for research-shaped work, and those uses vary by discipline. The next step is not to turn session counts into a productivity headline. It is to test specific workflows against shared scientific standards, with outcomes that another lab can reproduce.

We cover similar distinctions in our explainer on AI benchmarks and everyday agent tasks.