
A report from The Anthropic Institute, based on 207,049 research sessions, finds that scientists use Claude most where AI already performs well, not where they rank the hardest problems. The report says usage tracks a measure of theoretical AI capability across fields and tasks, rather than matching the tasks researchers describe as most important for discovery.
The report analyses 128,914 sessions on Claude (Chat and Cowork) and 78,135 on Claude Code, run between August 2 and September 30, 2026. A privacy-preserving classifier labelled each session by field, subfield and research task using the SciNet taxonomy: 7,262 tasks across 34 fields and 318 subfields. The report states that no open-ended research content was retained.
According to the report, the pattern holds across fields. Computer science and statistics use Claude far more than their share of researchers would predict; medicine and the lab-based physical sciences use it less. Data analysis accounts for 33.7% of Claude sessions and 48.1% of Claude Code sessions, against 27.6% of researcher time. Ideation accounts for 4.2% and 2.1%, against 7.1% of researcher time.
The report sets out three acceleration scenarios. It stresses that they are illustrative, not forecasts. Under wide adoption of current capabilities, research output could roughly double in the most AI-ready fields. Under 75% automation of non-physical steps, computer science could reach about 8x. Under full automation of non-physical steps, the ceiling is about 16x, set by the ratio between AI time and researcher time.
The full report is at anthropic.com/research/how-researchers-use-claude.
The report finds that usage tracks capability. Claude use is highest for tasks that score highest on the report’s measure of theoretical AI capability. The report says the top quintile of tasks by capability gets two to four times the use its share of researcher time would predict, while the bottom quintile, which includes most physical work, gets very little. The pattern is reported across 207,049 sessions and across fields.
The report identifies a gap between use and stated importance. Data analysis takes 27.6% of researcher time and 33.7% of Claude sessions. Writing takes 16.0% of time and 32.0% of sessions. Data gathering takes 28.6% of time and 11.5% of sessions. Ideation takes 7.1% of time and 4.2% of sessions. In the report’s accompanying survey, researchers rank ideation first and data gathering second as drivers of important results. Those are the two categories they use AI least for.
The report says field differences follow the same logic. Computer science, mathematics and statistics use Claude more than their researcher share predicts. Medicine and the lab-based physical sciences use it less. The report’s reading is that fields whose work is mostly cognitive and code-based adopt AI faster, while fields whose work is physical or regulated adopt it more slowly.
| Task category | Share of Claude sessions | Share of Claude Code sessions | Share of researcher time | Survey rank for importance |
|---|---|---|---|---|
| Data analysis | 33.7% | 48.1% | 27.6% | 3rd |
| Writing and communication | 32.0% | 17.6% | 16.0% | Not ranked in top 4 |
| Data gathering | 11.5% | 11.9% | 28.6% | 2nd |
| Reading and knowledge acquisition | 7.6% | 3.7% | — | — |
| Theoretical analysis | 5.0% | 4.7% | — | — |
| Design and development | 4.7% | 7.7% | — | — |
| Ideation and hypothesis generation | 4.2% | 2.1% | 7.1% | 1st |
| Other (admin, service) | 1.3% | 4.2% | — | — |
The two categories researchers rank highest are the two they use AI least for. The report offers three explanations, each consistent with the data. Data gathering is often physical: collecting clinical or biochemical data requires hands, instruments or patients. Ideation may be work researchers prefer to do themselves, or AI may be weaker at high-level ideation, such as choosing which research question to pursue and designing a study to answer it.
The report says specialised tools narrow the gap. On Claude Science, which connects Claude to research databases and tools, ideation makes up 8% of sessions, roughly the share of researcher time. The report says that is about twice the ideation share on Claude and four times the share on Claude Code.
| Rank | Field | Share of Claude sessions | Share of Claude Code sessions | Share of world researchers |
|---|---|---|---|---|
| 1 | Computer science | 11.9% | 32.7% | 4.3% |
| 2 | Engineering | 8.4% | — | — |
| 3 | Medicine | 7.7% | — | — |
| — | Mathematics and statistics | — | High relative use | — |
Key point: The report notes that the two largest fields by session share on Claude are also among the largest by global researcher counts. The gap opens in the ratio between the two. Computer science is 4.3% of the world’s researchers and 32.7% of Claude Code sessions, according to the report. That ratio is the pattern the report spends most of its length explaining.
Two-step classification. The report describes a first classifier that identifies whether a session involves scientific work, and a second that assigns each scientific session to a field, subfield and task in the SciNet taxonomy. The only information retained from each session is the resulting labels, such as “chemistry” or “data cleaning.” The report states that no open-ended research content is extracted or stored.
The taxonomy. The report uses SciNet, which contains 7,262 tasks across 34 fields and 318 subfields. Some tasks, such as drafting manuscripts, are universal. Others, such as predicting optical absorption and emission spectra, are specific to one subfield. Together the tasks cover the full process of producing a new research finding.
Two samples. The report’s main sample is 207,049 sessions on consumer plans from August 2 to September 30, 2026: 128,914 from Chat and Cowork (referred to as “Claude”) and 78,135 from Claude Code, an agentic coding tool. A supplemental sample covers Claude Science sessions. A survey of 528 scientists provides self-reported AI use for comparison.
The full report is at anthropic.com/research/how-researchers-use-claude. The appendix is linked separately.
| Sample | Size | Period | What it covers |
|---|---|---|---|
| Claude (Chat and Cowork) | 128,914 sessions | Aug 2 – Sep 30, 2026 | General-purpose conversation and knowledge work |
| Claude Code | 78,135 sessions | Aug 2 – Sep 30, 2026 | Agentic coding with access to a user’s files |
| Main sample total | 207,049 sessions | Aug 2 – Sep 30, 2026 | Sessions assigned a field and a task |
| Claude Science | Supplemental sample | Aug 2 – Sep 30, 2026 | Sessions on the specialised science interface |
| Researcher survey | 528 respondents | 2026 | Life and health sciences, active AI adopters |
| Dimension | Count | Example |
|---|---|---|
| Fields | 34 | Computer science, chemistry, medicine |
| Subfields | 318 | Biomedical engineering, econometrics |
| Tasks | 7,262 | Drafting manuscripts; purifying reaction mixtures |
| Task categories | 8 | Data analysis; writing; data gathering; ideation |
How capability scores are built. Each task in SciNet lists the individual steps a scientist must perform to complete it. For example, one chemistry task involves purifying reaction mixtures, and one of its steps involves packing a chromatography column. The report describes a classifier shown pairs of steps and asked which one AI can perform better. A Bradley-Terry model, the same model behind Elo ratings, aggregates the comparisons into a step-level score. Each task’s score is the average of its steps’ scores. The report calls the result “theoretical AI capability.”
How the survey was used. The report says the 528 respondents were recruited from applicants to Anthropic’s AI for Science and related grant programs, and are affiliated with research institutions in the US and Canada. The report notes that because they applied for AI research credits, they are likely more intensive AI users than the typical scientist. The report treats their self-reported use as a proxy for what AI can do today, and flags the selection issue in the appendix.
Privacy safeguards. The report states that results are reported only in aggregate and only for cells that meet a minimum number of distinct accounts. No classifier in the study stores research content, datasets, hypotheses, experimental methods, early findings or unpublished papers. The classifiers are listed in the appendix.
Session counts by field, subfield and task, assigned by classifier.
Self-reported AI use by task category, from the survey.
Artifacts produced per session, from a separate classifier.
Share of sessions vs share of researcher time, by category.
Use ratio: session share divided by researcher-time share.
Capability scores by task, correlated with usage and with task traits.
The report says field size explains a lot, not all. On Claude, the report reports a log-log correlation of 0.79 between a field’s share of researchers and its share of sessions. On Claude Code it reports 0.62. Fields above the 45-degree line in the report’s Figure 1 use Claude more than their size predicts. Those include computer science, mathematics and statistics. Fields below the line use it less. Those include chemistry and medicine.
The report says subfields sharpen the picture. Artificial intelligence and machine learning is 1.2% of the world’s researchers and 5.0% of Claude sessions and 15.2% of Claude Code sessions, according to the report. Computational subfields lead regardless of the parent field: computational biology, computational linguistics, computational neuroscience. Robotics also shows high relative use, the report says, because it builds on computational methods despite its focus on physical systems.
| Field group | Relative Claude use | Relative Claude Code use | Why, per the report |
|---|---|---|---|
| Formal sciences (CS, maths, stats) | High | Very high | Cognitive, code-based, high capability scores |
| Social sciences | High | Low except economics | Text-based; economics is comparatively quantitative |
| Life sciences | Moderate | Moderate | Some computational subfields lead |
| Physical sciences | Low | Low | Lab-based, physical bottlenecks |
| Health sciences (medicine) | Low | Low | Time with patients; regulatory limits |
| Arts and humanities | Low | Very low | Qualitative and interpretive work, lower capability scores |
| Subfield | Share of Claude sessions | Share of Claude Code sessions | Share of world researchers |
|---|---|---|---|
| AI and machine learning | 5.0% | 15.2% | 1.2% |
| Bioinformatics and computational biology | 3.1% | 6.5% | — |
| Computer vision | 1.5% | 4.1% | — |
| Epidemiologic methods | 1.9% | — | — |
| Software engineering | — | 2.7% | — |
| Robotics | — | 2.6% | — |
| Natural language processing | — | 2.2% | — |
| Econometrics | High relative use | High relative use | — |
| Task | Claude | Claude Code | Researcher time | Use ratio, Claude |
|---|---|---|---|---|
| Data analysis | 33.7% | 48.1% | 27.6% | 1.22 |
| Writing and communication | 32.0% | 17.6% | 16.0% | 2.00 |
| Data gathering | 11.5% | 11.9% | 28.6% | 0.40 |
| Ideation and hypothesis generation | 4.2% | 2.1% | 7.1% | 0.59 |
| Design and development | 4.7% | 7.7% | — | — |
| Reading and knowledge acquisition | 7.6% | 3.7% | — | — |
| Theoretical analysis | 5.0% | 4.7% | — | — |
The report says self-reports match the session data. The 528 surveyed scientists report their highest AI use for data analysis and writing, and their lowest for ideation and data gathering, according to the report. They also report more shared or delegated use of AI for data analysis and writing. The report says that lines up with the task shares observed in the sessions.
The report finds the importance ranking runs the other way. When asked to allocate 100 points across four categories for producing an important research result, respondents give ideation 35.6 points and data gathering 23.6, against 23.3 for data analysis and 17.5 for writing. The report notes that the categories they rate highest are the categories they use AI least for.
The report reports doctoral-level content on Claude Science. Over 80% of Claude Science sessions are estimated to require doctoral-level training (20 years of education), compared with about 48.0% on Claude and 67.7% on Claude Code. The report says sessions on Claude Science contain work that would take 10.3 human hours to complete without AI, against 4.9 hours on Claude and 5.9 on Claude Code. Over 70% of Claude Science sessions are performed with a high degree of AI autonomy, against 23.7% on Claude and 64.1% on Claude Code.
| Measure | Claude | Claude Code | Claude Science |
|---|---|---|---|
| Human hours per session | 4.9 | 5.9 | 10.3 |
| Doctoral-level content | 48.0% | 67.7% | 82.7% |
| High AI autonomy | 23.7% | 64.1% | 73.3% |
The report says use rises with the capability score. For tasks in the lowest quintile of capability scores, such as purifying reaction mixtures or conducting participant observation, Claude is barely used, according to the report. At the median, the report says the use ratio is about 1, meaning use matches researcher time. For the top quintile, such as visualizing data or applying statistical tests, the report says the ratio is about 2x on Claude and about 4x on Claude Code.
The report finds capability correlates with task traits. The report reports the capability score correlates at 0.78 with an RL-suitability index, at 0.69 with cognitive work, and negatively with physical work (-0.57), interpersonal work (-0.31) and regulatory restriction (-0.38). RL stands for reinforcement learning: the model tries a task repeatedly and is rewarded when it succeeds. Tasks are suitable for RL when success is easy to verify, many examples can be simulated, and feedback is fast.
The report says subfield-level patterns confirm it. A subfield’s Claude use per researcher rises with the average capability score of its tasks, according to the report. The reported correlation is 0.45 for Claude and 0.57 for Claude Code. The report also finds that autonomous use correlates with capability: 0.47 on Claude and 0.31 on Claude Code.
| Task trait | Correlation with capability score | Direction of Claude use, per the report |
|---|---|---|
| Reinforcement-learning suitability | +0.78 | Rises sharply with RL suitability |
| Cognitive work | +0.69 | Rises with cognitive intensity |
| Physical work | -0.57 | Falls with physical intensity |
| Interpersonal work | -0.31 | Falls with interpersonal intensity |
| Regulatory restriction | -0.38 | Falls where rules limit AI use |
The capability gradient. The report says the use ratio rises monotonically with capability quintile, and the same pattern shows up across 318 subfields.
The task-share mismatch. The report says data analysis and writing are overrepresented relative to researcher time; data gathering and ideation are underrepresented. The report says the survey confirms both.
The report says the survey is small and selected: 528 respondents, US and Canada, life and health sciences, AI-grant applicants.
The report says Claude Science adopters are self-selected, so the product comparison describes patterns of use, not the effect of the tool.
The report does not measure enforcement or diffusion. The sessions are a snapshot of active users, not a random sample of scientists.
Two measures per scenario. The report defines the AI share as the fraction of researcher time spent on a study that AI could take on, given how researchers currently allocate their time. It defines the acceleration as the increase in research output for a fixed research budget. The report notes that freed-up time does not translate one for one into more output, because researchers redirect effort toward the tasks where AI helps less, and because AI itself has costs.
The production function. The report uses an ideas production function in which a paper requires many complementary research tasks. The report says strength in one cannot fully make up for weakness in another. As AI makes some tasks cheaper, researchers shift effort toward tasks where AI helps less, easing the bottleneck. The report calibrates the bottleneck using Ahmadpoor and Jones (2019).
The report’s cost assumptions. The report assumes an hour of AI agent time costs $30, against $95 for an hour of researcher time, and that AI runs for 0.2 hours for each researcher hour it saves. The report says fully automated tasks cost about one-sixteenth of what a researcher would cost, so acceleration approaches about 16x and stops there. The report adds that halving the cost of AI would double the rate of progress, under these assumptions.
| Scenario | What is automated | Assumption | Status, per the report |
|---|---|---|---|
| Scenario 1: Current capabilities | Steps at the levels reported in the survey: none, assistive, shared, delegated, autonomous | Wide adoption of existing AI capabilities | Grounded in the survey |
| Scenario 2: 75% automation | 75% of non-physical research steps, starting with the highest capability scores | AI handles three-quarters of cognitive work | Hypothetical |
| Scenario 3: 100% automation | All non-physical research steps | AI handles all cognitive work; physical steps remain human | Hypothetical |
| Field | Scenario 1: current | Scenario 2: 75% | Scenario 3: 100% |
|---|---|---|---|
| Statistics | 2.02x | 7x | ~15-16x |
| Computer science | 1.96x | 8x | ~15-16x |
| Economics | High | 6x | ~15-16x |
| Library and information science | High | 7x | ~15-16x |
| Mathematics | High | 4-5x | ~15-16x |
| Physics and astronomy | Moderate | 4-5x | ~15-16x where cognitive |
| Psychology | Moderate | 4-5x | ~15-16x |
| Chemistry | 1.41x | 2-3x | 3-4x |
| Biology | Moderate | 2-3x | 3-4x |
| Biomedical sciences | Moderate | 2-3x | 3-4x |
| Materials science | Moderate | 2-3x | 3-4x |
| Agricultural sciences | Moderate | 2-3x | 3-4x |
| Arts | 1.35x | ~2x or less | ~15-16x in Scenario 3 |
| Anthropology | 1.36x | ~2x or less | ~15-16x in Scenario 3 |
| History | Low | ~2x or less | ~15-16x in Scenario 3 |
The report says physical bottlenecks cap the gains. In Scenario 1, the report finds AI could take on 27-52% of the researcher time needed to produce a new study, with a median of 38% across fields. In lab and field sciences, the report says 38-41% of researcher time goes to physical steps, against 4% in the median field. The report says the physical share alone explains about one-third of the variation in the AI share.
The report says capability scores explain most of the rest. The report notes that the arts and history spend little time on physical steps (10% and 4%) yet gain little, because their cognitive steps have low capability scores. Computer science and statistics gain most, the report says, driven by data analysis.
The report says the ceiling moves with cost. In Scenario 3, fields with few physical tasks reach about 15-16x, according to the report. Fields with substantial lab or field work reach only 3-4x. The report says lower AI costs would raise the ceiling proportionally.
| Task | Already possible today | By 2030 |
|---|---|---|
| Data analysis | 38% | 94% |
| Writing and communication | High | 89% |
| Ideation and hypothesis generation | Low | 72% |
| Data gathering | Low | 72% |
| Full research cycle (no human in the loop) | Low | ~40% |
| Breakthrough finding (no human in the loop) | Very low | About half expect only after 2035, if ever |
The report says forecasters have underestimated AI progress. The report cites a mid-2022 survey that put the median estimate for AI achieving gold-level performance in the International Mathematical Olympiad at 2030 among domain experts and 2035 among superforecasters. The report notes that AI achieved the milestone in 2025, five to 10 years ahead of those predictions. The report cites this as evidence that progress is often faster than expected, with exceptions in domains that require physical execution.
Researchers. The report’s data show where AI is already useful and where it is not. The report says the strongest uses are data analysis, writing and code-based work. The report says the weakest are physical data collection and high-level ideation. A researcher deciding where to deploy AI can read the task shares as a map of where returns are already visible, according to the report.
Research leaders. The report says physical bottlenecks cap AI’s contribution in lab and field sciences. The report points to investment in automated labs and digital data collection as where the physical share can be reduced. The report treats this as one of the conditions that would move a field from the 3-4x ceiling toward the 15-16x ceiling.
| Group | What they can do | What the report’s data support |
|---|---|---|
| Individual researchers | Focus AI use on data analysis, writing and code; keep ideation and physical data collection human | Task shares and capability scores, pp. 9-17 |
| Lab and field scientists | Invest in automating physical steps to lower the bottleneck | Physical share explains one-third of the AI-share variation, p. 25 |
| Research institutions | Support adoption in formal and computational subfields, where returns are largest | Subfield use data, p. 7 |
| Funders | Weigh proposals for automated labs and data infrastructure | Scenarios 2 and 3, pp. 26-27 |
| Policymakers | Account for regulatory restrictions that lower AI use in health and biomedical research | Regulatory correlation of -0.38, p. 18 |
| AI developers | Target the tasks where capability scores are low but researcher time is high, especially ideation and data gathering | Use-importance gap, pp. 9-11 |
Researchers could redirect AI use toward the tasks where the report’s capability gradient is steepest.
Research offices could plan for the productivity gains in formal and computational fields, and for the slower gains in lab sciences.
Funders could see where physical infrastructure, not AI capability, is the binding constraint.
BSD’s reading, not a claim the report’s authors make.
Scientists outside the sampled interfaces, or outside the US and Canada, where the report’s survey was drawn.
Researchers in fields where the physical share is far above the sample median.
Anyone reading the report’s scenarios as forecasts. The report stresses they are illustrative.
The report says the scenarios are illustrative, not forecasts. The report says they depend heavily on assumptions about AI’s reach and costs. The report’s cost assumptions ($30 AI hour, $95 researcher hour, 0.2 AI hours per researcher hour saved) are taken from today’s public list prices. The report notes that different assumptions move the results.
The report says several constraints are not counted. Field-level limits, such as how many experiments a facility like the Large Hadron Collider can run, are not in the model. Machinery and input costs for experiments are not counted, which matters most in health, physical and life sciences. Simulations that could replace physical experiments are not modelled. And the quality of AI-generated ideas is not modelled, in either direction.
The report says adoption lags capability. The report cites the diffusion literature: new technologies spread slowly, and productivity gains materialise only after organisations restructure around them. The report says individual learning, organisational friction and access constraints all slow the path from capability to output.
| Limit | Why it matters, per the report | Where it bites |
|---|---|---|
| Adoption lag | Capability does not translate automatically into use | All fields, especially slow-moving institutions |
| Physical bottlenecks | About 40% of researcher time in lab sciences is physical | Chemistry, biology, materials, agricultural, biomedical |
| Ideation quality | If AI-generated hypotheses are weaker, gains fall; if stronger, gains rise | All fields, hardest to measure |
| Quantity vs quality | A surge in papers could overwhelm absorption or raise the bar | Journals, funders, institutions |
| Organisational change | AI could take over verification, refereeing or communication | Peer review, research communities |
| Safety and alignment | More autonomous AI research raises oversight questions | Biology especially, where dual-use risk is high |
| Job satisfaction | AI could take over tedious tasks or rewarding ones; the net effect is uncertain | Research careers |
| Not measured in the report | What would change |
|---|---|
| Enforcement and compliance | The report says actual use may differ from what the sessions show if disclosure or rules vary by institution |
| Random sample of scientists | The report says session data come from active users, not a random sample; the survey skews toward AI-grant applicants |
| Automated labs | If physical steps become automated, the bench-science ceiling rises toward the cognitive ceiling |
| AI cost declines | The report says lower AI cost raises the ceiling proportionally; halving cost doubles the rate of progress |
| Idea quality | The report says better AI hypotheses accelerate further; uniquely human research taste slows it |
Three open questions the report leaves. First, whether specialised science tools such as Claude Science can close the ideation gap at scale. Second, whether automated labs can reduce the physical bottleneck in bench sciences. Third, whether the survey’s respondents, who expect about 40% of full research projects to be AI-run by 2030, will turn out to be right or, as with past AI forecasts cited in the report, too conservative.
The full report and its appendix are at anthropic.com/research/how-researchers-use-claude.
Scenarios are illustrative, not forecasts.
Field-level constraints and machinery costs are not counted.
Idea quality and simulation substitution are not modelled.
The session sample is active users on consumer plans, not a random sample of scientists.
The survey is 528 respondents, US and Canada, life and health sciences, AI-grant applicants.
Claude Science adopters are self-selected, so the product comparison describes patterns of use, not the effect of the tool.
What comes next: The report suggests a follow-up could sample scientists outside the grant-applicant pool and track whether the ideation gap closes as specialised science tools mature.