Anthropic study of 207,049 Claude sessions shows how scientists use AI
Anthropic published a study on November 5, 2026 that measures how researchers actually use Claude, based on 207,049 research sessions on consumer plans recorded between August 2 and September 30, 2026. The sample splits into 128,914 sessions in Claude (the chat app and Cowork) and 78,135 sessions in Claude Code, with a smaller supplemental sample from Claude Science, the company's research workbench that connects Claude to scientific tools and databases. The work was carried out by Anthropic's Economics team, with Johannes Hermle, Ben Jones and Lukas Althoff as lead authors.
The headline result concerns tasks rather than fields. Data analysis accounts for 33.7% of research sessions in Claude and 48.1% in Claude Code, against an estimated 27.6% of the time researchers spend on that category. Writing and communication take 32.0% of Claude sessions, twice its 16.0% share of researcher time. Ideation and data gathering, which the scientists Anthropic surveyed rank as the most important parts of discovery, are used far less than their weight in a researcher's working week would suggest.
Anthropic also models what wider adoption could mean for research output. In its first scenario, which assumes only current capabilities used broadly, AI could take over between 27% and 52% of research time depending on the field, with a median of 38%, and research output would rise by a factor ranging from 1.35x in the arts to 2.02x in statistics. The company presents these figures as illustrative scenarios and states plainly that they are not forecasts.
··········
HOW ANTHROPIC MEASURED CLAUDE RESEARCH SESSIONS.
The study relies on a privacy-preserving pipeline. A language-model classifier assigned each session a field, a subfield and a research task drawn from the SciNet taxonomy, which covers 7,262 tasks across 34 fields and 318 subfields. Only those labels were kept, the conversations themselves were not retained for the analysis, and results are reported in aggregate and only for cells that contain enough distinct accounts.
To judge whether a task is over- or under-represented, the authors compare each category's share of Claude sessions with its share of researcher time as estimated by SciNet. The resulting figure is called a use ratio: a value above 1 means scientists bring that kind of work to Claude more often than its weight in their schedule would predict, and a value below 1 means the opposite.
A third layer estimates how well AI could perform each task. A language model compared pairs of research steps and judged which one AI could carry out better, and a Bradley–Terry model turned those pairwise judgments into step scores, which were then averaged into task scores. Anthropic complemented the usage data with a survey of 528 scientists, mostly in the life and health sciences and affiliated with institutions in the United States and Canada, recruited from applicants to its AI for Science and related grant programs.
The sample has clear boundaries. It covers consumer plans only, and the study does not say whether that means Free, Pro, Max and Ultra together or a subset of them. Team, Enterprise and API traffic is not described as part of the sample, so the picture reflects individual researchers who use Claude on their own accounts. The survey population is also likely to be heavier on AI use than the average scientist, a limitation the authors acknowledge.
··········
WHICH RESEARCH TASKS SCIENTISTS BRING TO CLAUDE AND CLAUDE CODE.
The task breakdown is the most detailed part of the study. The table below lists the eight task categories Anthropic reports, with the share of sessions in each product and, where the study gives it, the share of researcher time.
........
Task category | Claude sessions | Claude Code sessions | Share of researcher time |
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% |
Reading and knowledge acquisition | 7.6% | 3.7% | Not stated |
Theoretical analysis | 5.0% | 4.7% | Not stated |
Design and development | 4.7% | 7.7% | Not stated |
Ideation and hypothesis generation | 4.2% | 2.1% | 7.1% |
Other | 1.3% | 4.2% | Not stated |
........
At the level of individual tasks, drafting manuscripts is the single largest use, at 18.6% of Claude sessions and 11.1% of Claude Code sessions. In Claude, reviewing and synthesizing literature follows with 6.8% and responding to peer reviewer comments with 6.7%. In Claude Code, designing, building and refining artifacts takes 4.8% of sessions and visualizing data and results 4.6%.
Dividing the session shares by the time shares for the four categories where Anthropic gives both values produces the use ratios below. The calculation is ours, based on the study's published percentages.
........
Task category | Use ratio in Claude | Use ratio in Claude Code |
Data analysis | 1.22 | 1.74 |
Writing and communication | 2.00 | 1.10 |
Data gathering | 0.40 | 0.42 |
Ideation and hypothesis generation | 0.59 | 0.30 |
........
The pattern is consistent across both products. Writing work in the chat app and analysis work in Claude Code run well above their share of a researcher's time, while data gathering sits at roughly 0.4 in both and ideation falls to 0.30 in Claude Code. The survey makes this gap more visible: when respondents divided 100 importance points across four categories, they gave 35.6 to ideation, 23.6 to data gathering, 23.3 to data analysis and 17.5 to writing. The tasks scientists consider most valuable for discovery are the ones they delegate least.
··········
CLAUDE USAGE BY SCIENTIFIC FIELD AND THE CLAUDE CODE SKEW.
In Claude, the largest field shares belong to computer science with 11.9% of sessions, engineering with 8.4% and medicine with 7.7%. The study finds that the formal sciences and the social sciences use Claude more than their share of the world's researchers would predict, while chemistry and medicine use it less.
Claude Code is far more concentrated. Computer science accounts for 32.7% of its research sessions while representing 4.3% of researchers, about 7.6 times its expected weight, compared with roughly 2.8 times in the chat app. The subfield of AI and machine learning is even more skewed: it represents 1.2% of researchers but 5.0% of Claude sessions and 15.2% of Claude Code sessions, ratios of about 4.2 and 12.7.
Anthropic quantifies how closely usage follows the size of each discipline with log-log correlations between session share and researcher share. The values are 0.79 for fields in Claude, 0.62 for fields in Claude Code, 0.61 for subfields in Claude and 0.43 for subfields in Claude Code. Usage in the chat app therefore tracks the global distribution of researchers fairly closely, while the coding tool is pulled toward disciplines where software is already the main instrument.
The outputs differ by discipline too. Across research sessions in Claude, 52% produced written artifacts and 21% produced technical artifacts such as code or data files. Biology was the only field in which technical artifacts outnumbered written ones.
··········
CLAUDE SCIENCE SESSIONS IN BIOLOGY AND BIOMEDICAL RESEARCH.
The supplemental sample from Claude Science looks different from both main products. Biomedical sciences account for 31.1% of its sessions and biology for 24.1%, with medicine and clinical sciences at 8.7% and every other field below 5%. Ideation reaches about 8% of Claude Science sessions, roughly in line with its share of researcher time, which makes it the only one of the three products where hypothesis work is not under-represented.
The task mix inside Claude Science is also more specialized. Proposing mechanistic models for protein–protein or protein–ligand interactions is 5.6 times more likely there than in Claude. Anthropic's estimates of session depth show the same gradient across the three products.
........
Metric | Claude | Claude Code | Claude Science |
Human time equivalent per session | 4.9 hours | 5.9 hours | 10.3 hours |
Sessions with doctoral-level content | 48.0% | 67.7% | 82.7% |
Sessions with high AI autonomy | 23.7% | 64.1% | 73.3% |
........
By these estimates, an average Claude Science session corresponds to about 2.1 times the human work of an average chat session. The authors caution that part of the difference may come from who chooses to adopt Claude Science, since a tool built for specialists attracts specialists, rather than from the tool alone.
··········
WHAT THE CLAUDE STUDY MEANS FOR RESEARCHERS ON PAID PLANS.
For working scientists, the practical information starts with access. According to Anthropic, any paid Claude user can open Claude Science; the study does not give separate pricing, usage limits or regional details for the workbench. Readers comparing what each paid tier includes can check our guide to Claude plans and usage limits on Pro, Max, Ultra, Team and Enterprise. Researchers who want credits rather than a subscription can look at Anthropic's AI for Science program, which supplied the survey sample; the study itself does not publish the program's terms, amounts or eligibility rules.
The usage data also suggests where scientists are getting the most from the current models. Data analysis is the dominant task in both products, and Claude Code users hand off more of it, with 64.1% of Claude Code research sessions classified as high-autonomy against 23.7% in the chat app. Groups that analyze large datasets, maintain analysis code or produce figures are the ones already relying most heavily on agentic work, and the coding tool is where that work concentrates.
The gap on ideation and data gathering points the other way. These two categories take 35.7% of researcher time in SciNet's estimates but only 15.7% of Claude sessions, even though survey respondents rate them as the most important steps. Respondents' own judgment of current ability matches this: 38% say AI can already handle data analysis, 30% writing, 23% data gathering and 13% ideation.
Expectations for the next few years are high among this group. By 2030, 94% expect AI to match top researchers in data analysis, 89% in writing, and 72% each in ideation and data gathering. About 40% expect AI to carry out full research projects by 2030, about 64% by 2035, and 16% say it will never happen. For related context on how another lab frames research automation, see our report on OpenAI's automated research intern and its research acceleration claims.
··········
ANTHROPIC'S RESEARCH ACCELERATION SCENARIOS BY FIELD.
The acceleration estimates come from an ideas production function in which AI substitutes for part of researcher time. The model's economic assumptions are explicit: AI costs $30 per hour against $95 per hour for a researcher, and AI needs 0.2 hours of running time for each researcher hour it saves. On those numbers, each researcher hour saved costs about $6 in AI time, roughly 6.3% of the researcher's hourly cost.
........
Scenario | Assumption | Result reported by Anthropic |
1 | Current capabilities, widely adopted | 27–52% of research time taken on by AI; output up 1.35x (arts) to 2.02x (statistics); computer science 1.96x, chemistry 1.41x |
2 | 75% of non-physical research steps automated | Computer science about 8x, statistics about 7x, economics about 6x; bench and medical sciences about 2–3x |
3 | All non-physical research steps automated | Low-physical fields about 15–16x; biology, chemistry, materials, agricultural and biomedical sciences about 3–4x |
........
The spread between fields comes from physical work. In biology, chemistry, materials science, agricultural sciences and biomedical sciences, physical steps account for roughly 40% of researcher time, and the model assumes AI cannot shorten them, so even complete automation of everything else caps acceleration at about 3–4x. Capability scores line up with this: task capability correlates at 0.78 with suitability for reinforcement learning and 0.69 with cognitive work, and negatively with physical work (−0.57), interpersonal work (−0.31) and regulatory restrictions (−0.38).
Usage follows capability. Tasks in the lowest-capability fifth have a use ratio of about 0.2, tasks in the middle about 1, and tasks in the top fifth about 2.1 in Claude and 3.9 in Claude Code. Across subfields, capability correlates with use per researcher at 0.45 in Claude and 0.57 in Claude Code.
··········
CLAUDE RESEARCH USAGE STUDY: CONFIRMED FACTS AND OPEN QUESTIONS.
The measured part of the study is solid within its scope: 207,049 classified research sessions from consumer plans over two months, with a stable picture in which data analysis and writing dominate, Claude Code concentrates in computer science and AI research, and ideation and data gathering remain under-used relative to their place in scientific work. Claude Science, available to paid users, shows deeper and more autonomous sessions in the life sciences, with a selection effect the authors themselves flag.
Several elements remain uncertain. The classification depends on a language model's labels, the SciNet time benchmarks are estimates, the capability scores come from model judgments rather than lab results, and the survey draws on grant applicants who are likely heavier AI users than their peers. The acceleration scenarios leave out equipment and input costs, facility constraints at field level and the quality of AI-generated ideas, which is the variable most likely to decide whether higher output translates into more discoveries.
The figure most likely to be quoted, research output doubling in some fields, rests on the first scenario and applies at the upper end to statistics and computer science, at 2.02x and 1.96x. For chemistry the same scenario gives 1.41x, and for the arts 1.35x, which is the more representative range for disciplines where much of the work still happens at the bench, in archives or in the field.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
