AI for scientific discovery: How researchers use Claude

A November 2026 Anthropic Institute report that classifies 207,049 research sessions on Claude and Claude Code from August and September 2026 against the SciNet taxonomy of research tasks, finding heavier use in the formal sciences and for data analysis and writing, lighter use for data gathering and ideation, and use that rises with a model-judged score of AI capability on each task; it pairs this with a survey of 528 AI-using scientists and with illustrative scenarios in which current AI could up to double research output in some fields and full automation of non-physical work could raise it about 15-fold, capped by physical work and AI cost.

Johannes Hermle, Ben Jones, Lukas Althoff et al. · 2026-11-05 · original

Where in science is Claude actually used, and how much could AI speed up research if that use spread? A report published on 5 November 2026, whose page carries the browser title "How scientists use Claude" and is titled "AI for scientific discovery: How researchers use Claude" on the page, in the 35-page PDF and in the suggested citation, answers in three parts. It measures which fields and research tasks appear in a sample of Claude and Claude Code sessions; it tests whether use follows a model-judged score of how well AI can do each task; and it builds scenarios for research acceleration from those measures and a survey of scientists. The citation names the institution as The Anthropic Institute and the type as a technical report; the page labels the team Economics. The byline lists nine authors, with Johannes Hermle, Ben Jones and Lukas Althoff as lead authors, followed by Ryan Heller, Thomas Lyttelton, Carolyn Stein, Nathan Wilmers, Maxim Massenkoff and Peter McCrory; no affiliations are given. The version of record is the web report with its PDF and a 45-page appendix PDF; the numbers checked here agree between the web page and the PDF.

What the report does

Usage sample

The main sample is 207,049 research sessions from consumer plans (Free, Pro and Max) between 2 August and 30 September 2026: 128,914 from Chat and Cowork, which the report calls "Claude", and 78,135 from interactive Claude Code sessions ("How we conducted this research"; appendix A). The appendix describes three steps:

  • sample a fixed number of transcripts per hour and product, later reweighted by inverse sampling probability;
  • screen for scientific or research work, which leaves 148,734 Claude and 86,884 Claude Code transcripts;
  • assign each to a subfield and task, which succeeds for 94%, then keep only core research tasks.

Non-research tasks such as teaching and service account for 7.7% of scientific sessions on Claude and 4.2% on Claude Code (appendix A). A supplemental sample covers Claude Science, a product that connects Claude to research tools and databases; the main text refers to the appendix for its construction, but the appendix gives no size for it, and its dates only in a sentence that covers "the sample" as a whole.

The SciNet taxonomy and researcher time

Sessions are mapped to SciNet, a dataset by Cristian Alamos and Lukas Althoff (2026) licensed under CC BY 4.0, a taxonomy of 6 domains, 34 fields, 318 subfields, 7,262 tasks and 90,583 task steps (appendix B). SciNet was built with language models following O*NET's design. The appendix describes validation against about 100 papers per subfield, wet-lab protocols on protocols.io and human expert input. SciNet also estimates the hours researchers spend on each task, from model-estimated step times and per-researcher paper counts in OpenAlex. Its researcher counts come from classifying 746,517 sampled OpenAlex authors with Claude Haiku 5.5, which agreed with Claude Sonnet 5.5 on 85% of field and 74% of subfield labels in a 1,500-author check (appendix B, "Researcher counts").

Classifiers and their validation

Five classifiers, all run on Claude Sonnet 5, screen for scientific work, assign field and task, rate Claude's contribution, and (for Claude only) read a user's memory block for field and career stage (appendix C). Transcripts are truncated to 10,000 characters per turn and 100,000 in total. Further labels (autonomy, human task time, education level, artifacts) reuse classifiers from earlier Economic Index reports. The appendix prints every prompt.

There is no ground truth for real sessions, and the appendix ("Validation of classifiers") reports three checks:

  • Synthetic transcripts. On 5,000 transcripts generated by models from known subfield-task pairs, accuracy was 92% (domain), 83% (field), 74% (subfield), 69% (task) and 58% (task and subfield together); random assignment would score 0.3% on subfield. About half the synthetic transcripts still echoed the label's wording; excluding them lowers task accuracy to 64% and task-and-subfield to 54%.
  • Memory as a second signal. For sessions with a memory block, field shares from transcripts and from memory have a log-log correlation of 0.92 (Figure A1).
  • Restricted samples. Main results hold in the 60 to 62% of usage the classifier labels with high confidence, and in the 75 to 83% where Claude's contribution is rated at least co-producing; task-category shares move by at most 1.5 and 2.1 percentage points.

The appendix reports no check of the labels against human coding of real sessions.

Theoretical AI capability

The report's central explanatory variable is an AI capability score for each task, which it calls "theoretical AI capability". A judge model is shown two SciNet steps and picks the one for which AI's typical unaided output would come closer to what an experienced researcher would "accept without substantial correction". AI is defined as a computer-based agent with the best public AI systems, written communication only and no lab integration. A Bradley-Terry model (a model that turns pairwise wins into a rating, as Elo ratings do for chess) fitted to 317,560 comparisons ranks the steps. The 16% of steps judged impossible for AI in both of their first two comparisons are placed at the bottom. A task's score is the percentile of its steps' mean percentile (appendix B). The appendix does not name the judge model. It reports that the score correlates with the exposure labels of Eloundou et al. (2024) applied to SciNet, without giving the correlation. It reports no comparison with measured AI performance on the steps: the score is a model's judgment of what AI can do.

Other task characteristics come from SciNet or further model labels: an index of suitability for reinforcement learning (RL), the principal component of model ratings of verifiability, simulatability and feedback density, which explains 79.5% of their variance; the shares of cognitive, physical and interpersonal steps; and a 0-to-4 regulatory-restriction score (appendix B).

The survey

In September 2026 the authors emailed 1,813 researchers who had applied to one of several Anthropic token grant programs between May 2025 and August 2026; 528 completed the survey, a response rate of about 29% (appendix D; this note's arithmetic). Respondents are mainly in the biological and medical sciences, lean computational, and are affiliated with institutions in the US and Canada (appendix D; footnote 7). The survey is the baseline for a study of AI's effects on research (appendix D). No weighting is reported. Respondents gave, for each of four task categories, the distribution of their AI use on their past few papers across five levels (none, assistive, shared, delegated, autonomous), and split 100 points among the categories by importance for an important result (Figure 4; appendix E).

What it finds

Use by field

On Claude, the largest fields by session share are computer science (11.9%), engineering (8.4%) and medicine (7.7%); a field's session share and its share of researchers have a log-log correlation of 0.79 ("How Claude use varies across scientific fields"; Figure 1). Relative to field size, use is high in the formal sciences and the social sciences and low in lab-based physical sciences and medicine. On Claude Code the correlation is 0.62, and computer science, with 4.3% of the world's researchers, accounts for 32.7% of sessions. Footnote 3 notes that sessions are assigned to fields by the research pursued, not the tool. At the subfield level, artificial intelligence and machine learning holds 5.0% of Claude and 15.2% of Claude Code sessions against 1.2% of researchers (Figure 2). Footnote 4 warns that low-volume subfields are estimated less precisely.

Use by task category

Use is compared with SciNet's estimate of researcher time ("How Claude use varies across research tasks"; Figure 3):

Task category Claude Claude Code 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%
Ideation and hypothesis generation 4.2% 2.1% 7.1%

The figure's description gives the remaining categories, among them reading and knowledge acquisition at 7.6% and 3.7%. Drafting manuscripts is the most frequent single task (18.6% on Claude, 11.1% on Claude Code; footnote 6). Data-gathering tasks with physical barriers are the most underrepresented. The authors give several possible reasons for low ideation use: scientists' preferences, perceived limits of AI, the difficulty of classifying ideation woven into other tasks, or genuinely weaker AI at choosing questions.

In the survey, self-reported use is highest for data analysis and writing (Figure 4). For ideation, 17% report no AI use, 53% assistive and 24% shared use; for data analysis, 20% assistive, 44% shared and 28% delegated. Respondents rank ideation as most important (35.6 of 100 points), then data gathering (23.6), data analysis (23.3) and writing (17.5).

Claude Science

Biomedical sciences (31.1%) and biology (24.1%) lead Claude Science sessions (Figure 5). Within those fields, specialized tasks are overrepresented relative to Claude: hypotheses on protein structure and function at 6.0 times, and mechanistic models of protein interactions and drug-target hypotheses at 5.6 times. Coordinating with collaborators is at 0.15 times. Ideation is 8% of Claude Science sessions, which the report reads as "roughly on par" with researcher time; the text does not say whether this share covers all Claude Science sessions or only those in biology and biomedicine. In biology and biomedicine, Claude Science sessions involve work estimated at 10.3 human hours against 4.9 (Claude) and 5.9 (Claude Code), 82.7% doctoral-level content against 48.0% and 67.7%, and high AI autonomy in 73.3% against 23.7% and 64.1% (Figure 6). Reweighting Claude and Claude Code to Claude Science's subfields and tasks gives 5.1 hours, 55.1% and 25.2% for Claude and 5.5 hours, 71.8% and 61.4% for Claude Code, so the gaps remain. All three measures are model-estimated. Footnote 8 cautions that adopters of Claude Science likely have more specialized or demanding needs, and the authors therefore "describe the differences across products as patterns of use". Claude Science is an Anthropic product, and the comparison is Anthropic's.

Use tracks the capability score

The use ratio is a task's share of sessions divided by its share of researcher time. In the lowest capability quintile it is about 0.2; at the median, about 1; in the top quintile about 2.1 on Claude and 3.9 on Claude Code (Figure 8; values as described in the figure's text). The capability score correlates with the RL index at 0.78 and with the cognitive share at 0.69, and negatively with the physical share (-0.57), interpersonal share (-0.31) and regulatory restriction (-0.38) (Figure 9). These are correlations between model-generated labels of the same SciNet steps. Use ratios rise with RL suitability and cognitive share and fall with the others (Figure 10).

Across subfields, use per researcher rises with average capability score (weighted correlations 0.45 on Claude, 0.57 on Claude Code; Figure 11), and at the field level 0.52 and 0.55 (appendix F). Health sciences use Claude less than their scores predict, which the authors suggest may reflect regulation; formal sciences, social sciences and the arts and humanities use it more. Subfields with higher scores also give Claude more autonomy (Figure 12). The authors summarize that "scientists appear to use AI where it works best" (Introduction). Since the capability score is a model's judgment, this is evidence that use tracks a model's assessment of AI capability, not measured performance.

Other descriptive results

Across research sessions on Claude, 52% produce written artifacts and 21% technical ones, with biology the only field where technical artifacts are more common (Figure 7). Among users whose memory states a career stage, doctoral students use Claude most for data analysis (34.4%) and professors for writing (44.0%) (appendix F).

The acceleration scenarios

The report states that these "are illustrative scenarios, not forecasts" ("Approach").

How they are built

  • Scenario 1 assumes wide adoption of current AI. Survey use levels are mapped to time savings of 0, 25, 50, 75 and 100%, and assigned to SciNet steps in the survey's proportions, highest-scoring steps first. Respondents' reported use is treated as a proxy for current capability because they are likely heavy adopters.
  • Scenarios 2 and 3 assume AI fully automates 75% or 100% of non-physical steps, highest scores first. Physical steps stay with researchers in all scenarios.
  • The AI share is the fraction of researcher time AI could take on.
  • The acceleration is the increase in output for a fixed budget. It comes from an ideas production function in which tasks are complements, following Jones (2026). The bottleneck parameter (ρ = -1) is calibrated from Ahmadpoor and Jones (2019).

The assumed costs are $95 per researcher hour (a median doctorate salary plus 28% benefits) and $30 per hour of AI agent time, "an illustrative value based on public API list prices for Claude Opus 5.5. It is not a measured average cost". AI is assumed to run 0.2 hours per researcher hour saved, and AI-assisted work is assumed to be of the same quality as researchers' work (appendix E, "Empirical implementation").

Results

  • Scenario 1. AI could take on 27 to 52% of researcher time, median 38%. Accelerations range from 1.35 times (arts), 1.36 (anthropology) and 1.41 (chemistry) to 1.96 (computer science) and 2.02 (statistics) (Figure 13). Lab and field sciences spend 38 to 41% of researcher time on physical steps, against 4% in the median field. The physical share explains about a third of the variation in AI share. Fields with higher Claude use ratios have larger AI shares (weighted correlation 0.58; Figure 14). The authors present this as further evidence that use tracks capability, since the AI share uses no Claude usage data. The two measures are not wholly independent, as this note reads the appendix: both use SciNet's model-estimated researcher time (the use ratio as its denominator, the AI share as its step weights), and the AI share also ranks steps by the capability score, which the use ratio does not use.
  • Scenario 2. Accelerations are about 8 times in computer science, 7 in statistics and library and information science, 6 in economics, 4 to 5 in most social sciences, mathematics and physics, and 2 to 3 in bench and medical sciences.
  • Scenario 3. Fields with few physical tasks reach about 15 to 16 times and lab-heavy fields 3 to 4 times. The ceiling of about 16 times comes from the cost assumption: with these parameters, a fully automated step costs about one-sixteenth of a researcher doing it ($30/$95 × 0.2 ≈ 0.063, this note's arithmetic). Halving AI's cost would double that ceiling.

The authors list limits that cut both ways: no field-level constraints such as shared facilities, no costs of equipment or materials, no simulation replacing experiments, and no change in the quality of ideas ("Scenarios 2 and 3"). The appendix notes that AI costs are held fixed across scenarios, which may be conservative if inference gets cheaper or optimistic if harder tasks need more compute.

Researchers' expectations

Survey respondents were asked when AI working alone would match the best researchers in their field (Figure 15). 38% said it already does so for data analysis, and by 2030 (cumulating the figure's printed segments), 94% expected it for data analysis, 89% for writing and 72% each for ideation and data gathering. About 40% expected AI to carry out a full research cycle without a human in the loop by 2030 and about 64% by 2035; 16% said never. For "a finding your field would regard as a major advance", 30% said later than 2035 and 18% never. The report cites forecasting work (Karger et al. 2023; Rosenberg et al. 2026) for the view that AI progress has often been underestimated, with exceptions in physical-world domains. These are expectations of a sample of grant applicants who use AI, not measurements.

Limitations

As stated in the report:

  • the usage sample covers consumer plans only, and interactive Claude Code only;
  • assigning one task per session misses work that spans tasks (appendix C);
  • ideation may be undercounted because it is hard to classify;
  • low-volume subfields are imprecise (footnote 4);
  • Claude Science users self-select (footnote 8);
  • the survey sample is selected for heavy AI use (footnote 7);
  • the scenarios depend heavily on assumptions about AI's reach and cost ("Scenarios 2 and 3").

Further limits, as this note reads the documents:

  • Model-made measures throughout. Field and task labels, researcher hours, the capability score, the RL index and the physical share are all produced or estimated by language models. The validations are against synthetic transcripts, memory blocks and model-built datasets, not human coding of real sessions or measured AI performance.
  • Shared inputs. Several reported correlations, including capability against RL suitability and AI share against use ratio, relate quantities built from the same SciNet steps by models.
  • Usage share is not adoption. Session shares describe Claude's consumer traffic, not what share of scientists use AI.
  • Developer's report on its own products. The data cannot be checked outside the company. The comparison of Claude Science with Claude, both Anthropic products, describes the specialized product's use as more demanding, with self-selection noted (footnote 8).

Research questions

  1. How well does the AI capability score predict measured AI performance? The score is a model's pairwise judgment. Running agents on a stratified sample of SciNet steps and grading their outputs, by human experts, would show how closely judged capability tracks performance.
  2. Would human coders agree with the field and task labels on real sessions? Validation used synthetic transcripts, where accuracy fell when label wording was removed. A human-coded sample of real, privacy-cleared research sessions would measure accuracy where it matters.
  3. Does Claude Science's more demanding use survive a comparison of the same users? Footnote 8 notes self-selection. Comparing the same accounts' sessions on Claude Science and on Claude, or users before and after adopting it, would separate tool from user.
  4. How sensitive are the acceleration estimates to the cost, time and quality assumptions? The appendix gives the formulas. Published sensitivity ranges for the AI-hour price, the 0.2 running-time ratio and ρ would show which conclusions are robust.
  5. Is ideation underused or undercounted? The report offers both readings. Multi-task labeling of sessions, which the appendix proposes, would show how much ideation is embedded in other tasks.
  6. Do the survey respondents' expectations predict later capabilities? The survey is a baseline. Comparing the 2027 and 2030 expectations with later evaluations of AI on the same task categories would test them.
  7. What does the planned study of AI's effects on research find? The survey is its baseline wave. Its outcomes, measured against the baseline, would add evidence on output that usage shares cannot give.

How it connects

  • Anthropic Economic Index report: Economic primitives (January 2026) introduced the autonomy, task-time and education measures the report applies to Claude Science (cited as Appel et al. 2026).
  • Anthropic Economic Index report: Cadences (June 2026) is cited (as Massenkoff et al. 2026) for the artifact classifier used in Figure 7.
  • The Anthropic Economic Index and Labor market impacts of AI also compare usage with a measure of what AI could theoretically do; the report's capability score is described as similar in aim to the exposure measure of Eloundou et al. (2024). This note makes the connection to those notes.
  • Agentic coding and persistent returns to expertise (June 2026) uses a similar classifier pipeline on all Claude Code sessions and finds occupation groups including scientists close to software users on coding success; three authors are shared. This note makes the connection.
  • Coding agents in the social sciences (May 2026) is a baseline survey of social scientists for a parallel study, with Lyttelton, Wilmers and Massenkoff as authors. The page lists it as a related article, which is not a citation; the connection is this note's.
  • The Science blog notes, among them the Riemann zeta zeros result and Formalizing Fermat's Last Theorem, are cited in the introduction as examples of AI in mathematics; they are individual case studies, where this report measures aggregate use.
  • Clio describes the privacy-preserving approach behind aggregate-only reporting of this kind; the report itself refers to its "privacy-preserving infrastructure" without naming a tool.

References