How Scientists Use Claude: 207,000 Sessions Studied

§ 01How Anthropic measured it#

A language model read each session and labelled it with a field and a research task from SciNet, a taxonomy of 7,262 tasks across 34 fields and 318 subfields. Only those labels were kept, and results are reported in aggregate, only for cells with a minimum number of distinct accounts. Of the 207,049 sessions, 128,914 came from Chat and Cowork (which the report calls Claude) and 78,135 from Claude Code. A smaller sample from Claude Science, Anthropic’s research tool for scientists, was studied separately.

§ 02Where scientists use Claude#

On Claude, the largest fields by session share are computer science (11.9%), engineering (8.4%) and medicine (7.7%). Size explains much of that, but not all. The formal sciences, computer science, mathematics and statistics, use Claude far more than their headcount predicts, while chemistry and medicine, where work happens in a lab or with patients, use it less.

Claude Code concentrates even harder. Computer science is 4.3% of the world’s researchers but 32.7% of Claude Code research sessions. The artificial intelligence and machine learning subfield alone is 1.2% of researchers, 5.0% of Claude sessions and 15.2% of Claude Code sessions.

§ 03What they use it for#

The gap is the story. In the survey, scientists ranked ideation and hypothesis generation as the most important category for producing an important research result, and data gathering second. Those are the two categories they use Claude for least. Anthropic offers reasons: much data gathering is physical, scientists may prefer to generate ideas themselves, ideation is hard to classify and may be undercounted, and AI may be genuinely weaker at choosing which question to pursue.

Specialized tools change the mix. On Claude Science, ideation is 8% of sessions, roughly in line with the time researchers spend on it. Over 80% of Claude Science use is estimated to need doctoral-level training, against about 50-70% on the other interfaces, and over 70% of its sessions run with a high degree of AI autonomy, against 23.7% on Claude and 64.1% on Claude Code.

§ 04Use follows capability#

Anthropic built a score for how well AI can, in theory, perform each research task. For tasks in the top 20% on that score, such as visualizing data and running statistical tests, Claude use is about twice what researcher time would predict on Claude and about four times on Claude Code. Tasks in the bottom 20%, like purifying reaction mixtures, barely get used at all.

The score tracks how easy a task is to train with reinforcement learning, where success is easy to check and feedback is fast: the correlation is 0.78. It rises with cognitive work (0.69) and falls with physical work (-0.57), interpersonal work (-0.31) and regulatory limits on AI (-0.38).

§ 05How much faster research could get#

The report then models what AI could do to research output for a fixed budget, in three scenarios. “These are illustrative scenarios, not forecasts.”

Two things cap the gains. Physical work is the first: in lab and field sciences such as chemistry and materials science, 38-41% of researcher time goes to physical steps, against 4% in the median field. “Absent automated labs, their physical steps, representing about 40% of researcher time, remain the bottleneck.” The second is cost. The model assumes an hour of AI agent time costs $30 against $95 for a researcher’s hour, with AI running 0.2 hours for each researcher hour it saves, based on today’s public list prices.

The survey points the same direction: about 40% of researchers who actively use AI “expect AI models to carry out full research projects on their own by 2030.” Anthropic also flags the risk side, from less human oversight of what gets researched to changes in what a research career feels like, and notes that the fields with the biggest gains include those that build AI itself.

§ 06Where CellCog fits#

CellCog is our product, an AI employee platform, and this study maps closely onto how our employees divide work with the people who hire them. An AI research assistant takes the cognitive, checkable steps the report says AI does best: reading the literature, running the analysis, building the charts and drafting the write-up, while keeping its own memory of the project between conversations. The question, the judgment and anything physical stay with you, which is exactly where the study says scientists keep them today. For how Anthropic is widening its own science work, see our record on Anthropic’s AI for Science expansion.

§ 07What is not established yet#

• Beyond consumer plans. The sample covers consumer plans only; enterprise and API research use is not in it.
• Ideation counts. Anthropic says ideation is woven into other tasks and may be undercounted.
• Classifier scores. Capability scores, time shares and task labels all come from language-model classifiers, validated by Anthropic but not by outside researchers yet.
• The scenarios. Anthropic calls them illustrative, and they depend heavily on assumptions about AI’s reach and cost.

§ 08Where to watch#

• Whether specialized tools like Claude Science keep pulling ideation share toward researcher time.
• Automated labs, the one change that would lift the physical cap in bench sciences.
• A repeat of this study on a later model, to see whether use keeps tracking capability.

§ 09Sources#

• Anthropic, How researchers use Claude, November 5, 2026, read 18:50 UTC.
• Anthropic, Economic Index, the usage research series this report sits beside.
• Anthropic, Claude Code, the agentic coding tool behind 78,135 of the sessions.
• Our record on Anthropic’s AI for Science expansion.