How are scientists using Claude? 207,049 sessions point to analysis and writing
Anthropic mapped 128,914 Claude and 78,135 interactive Claude Code sessions to research tasks. It finds heavy use in writing, analysis and computational fields, but the sample covers one provider's consumer plans and its productivity forecasts are modelled scenarios, not measured gains.
The short version
- What happened
- Anthropic analysed 207,049 core-research sessions from consumer plans between 2 August and 30 September 2026: 128,914 Claude and 78,135 interactive Claude Code sessions. Sessions are not unique researchers or completed projects.
- Who it affects
- Researchers can design checks around drafts, code and analysis rather than treating each session as a completed result.
- What remains unproven
- The report is authored by Anthropic, which has a commercial interest in Claude use, and is not described as peer reviewed.
Keep in mind
The report is authored by Anthropic, which has a commercial interest in Claude use, and is not described as peer reviewed.
The sample covers one provider's consumer-plan sessions over two months and excludes programmatic Claude Code.
Sessions are not unique people, projects, papers or successful outputs; no unique-researcher denominator is published.
Classification and many SciNet characteristics are partly model-derived and cannot be independently checked from private transcripts.
The 528 survey respondents came from 1,813 Anthropic grant applicants and likely overrepresent computational, AI-intensive researchers.
Acceleration estimates depend on strong cost, quality and automation assumptions and are not observed productivity effects.
Source analysis · 2 linked sources. Primary analysed source date: 5 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The answer: use is concentrated where work is digital, legible and easy to hand off
Anthropic's study finds scientific Claude use concentrated in writing, analysis, literature work and computational tasks, while physical, interpersonal and tightly regulated steps receive less. The core evidence covers 207,049 sessions on consumer plans between 2 August and 30 September 2026—128,914 on Claude Chat or Cowork and 78,135 interactive Claude Code sessions. A session is an interaction unit, not a person, paper, discovery or successfully completed task.
Manuscript drafting was the largest named activity, accounting for 18.6% of Claude sessions and 11.1% of Claude Code sessions. Literature review represented 6.8% of Claude sessions and responding to reviewers 6.7%. Claude Code showed more building and visualisation: 4.8% of sessions mapped to building research artefacts and 4.6% to visualisation. Across analysed sessions, 52% produced a written artefact and 21% a technical artefact, according to Anthropic's classifiers.
The study replaces anecdotes with a large, privacy-preserved view of one product. It does not show that work was accurate, novel, accepted or faster, and it cannot represent researchers using other models, institutional accounts or no AI. Anthropic authors and sells Claude, creating a direct commercial conflict. The report is not stated as peer reviewed and belongs in Latest rather than Breaking.[1][2]
How the 207,049-session denominator was built
The study sampled session transcripts from Free, Pro and Max consumer accounts. For Claude Code it included interactive human use and excluded programmatic activity. A first language-model classifier identified 148,734 scientific Claude sessions and 86,884 scientific Claude Code sessions. A second classifier successfully assigned a subfield and task to 94%. The main analysis retained core tasks intended to generate findings or papers, yielding the 128,914 and 78,135 counts.
Anthropic sampled a constant number of sessions per hour and product rather than a constant share. It weighted each session by the inverse of its sampling probability: total sessions for that product and hour divided by the number sampled. This addresses changing sampling fractions but cannot create a representative census of scientists. One prolific user may contribute many sessions, and the minimum-account privacy rule is not a published unique-user denominator.
The privacy pipeline showed transcripts only to language-model classifiers, retained closed-list labels and reported cells only above a minimum number of accounts. That limits exposure of scientific material but prevents independent readers from checking classifications. Anthropic says a stricter high-confidence sample produced similar field patterns and a high-contribution subsample retained 75% of Claude research use and 83% of Claude Code use.[2]
The distinction that matters
From product sessions to productivity scenarios
The evidence becomes more assumption-dependent at each stage.
235,618 scientific sessions
Assign a field and task to 94%
Core sample selected
207,049 core sessions
Weight by product and hour
Usage shares
528 survey completions
Map AI roles to assumed time savings
Estimated time freed
Economic model
Add cost, quality and automation assumptions
Illustrative scenarios
SciNet turns conversations into fields, tasks and estimates
Anthropic used SciNet, a taxonomy covering six domains, 34 fields, 318 subfields, 7,262 tasks and 90,583 task steps. SciNet combines OpenAlex topics, language-model proposals, human input, papers, wet-lab protocols and O*NET-style descriptions. It estimates task frequency, researcher time and physical, cognitive, interpersonal and regulatory demands. The breadth is useful, but many labels and characteristics are model-derived rather than observed.
Scientific field use broadly tracked the global researcher population: the correlation between log usage per researcher and field size was 0.79. Computer science accounted for 11.9% of scientific sessions, engineering 8.4% and medicine 7.7%. These are shares among sampled sessions, not adoption rates within fields. Without a unique-user and eligible-population denominator, they cannot tell us what share of computer scientists or clinicians uses Claude.
Tasks in the highest quintile of theoretical AI capability received about 2.1 times their share of researcher time on Claude and 3.9 times on Claude Code; the lowest quintile received about 0.2 times. Reinforcement-learning suitability correlated 0.78 with capability, while physical demands correlated minus 0.57 and regulatory restriction minus 0.38. These are associations between constructed characteristics and session allocation, not causal evidence.[1][2]
A selected survey adds expectations, not a representative forecast
Anthropic surveyed researchers who had applied to its token-grant programmes between May 2025 and August 2026. In September it emailed 1,813 applicants and received 528 completed surveys, about 29%. The appendix says respondents came mainly from biological and medical sciences, leaned towards computational work and were likely stronger AI users because they selected into a programme offering model access.
The report says about 40% of surveyed active AI researchers expected AI to conduct complete projects by 2030. That is a belief among a selected group, not expert consensus or a measured forecast. Nonrespondents may differ, and Anthropic programme applicants are unusually connected to the provider. Survey reports of none, assistive, shared, delegated and autonomous use were mapped to 0%, 25%, 50%, 75% and 100% time savings.
Those mappings are assumptions, not time-and-motion measurements. A delegated draft can still demand source checking, code repair and experimental repetition. AI might also improve quality without saving time. The report acknowledges that its time-freed measure excludes AI costs and running time; later scenarios add assumed compute costs but still depend on translating qualitative roles into percentages.[1][2]
Productivity figures are scenarios, not measured scientific output
Anthropic's modelling suggests widespread use at current capabilities could roughly double output in some fields, while hypothetical full automation of non-physical steps could yield much larger gains—up to around 15 times in computer science and statistics and three to four times in laboratory sciences. These are not observed gains. They come from a production model combining SciNet time shares, survey-derived savings and assumptions about complementarity, running time, wages, AI costs and quality.
The appendix assumes an AI agent runs for 0.2 hours for every researcher-hour saved, costs $30 per hour and produces equal-quality work on the relevant step. It assumes a researcher-hour costs $95 and fixes a complementarity parameter from earlier literature. Future scenarios automate 75% or 100% of non-physical steps, beginning with the highest capability scores; physical steps are not newly automated.
The transparent assumptions are useful to interrogate but are not validated forecasts. Research quality, replication, equipment, ethics review, data access, peer review and the supply of valuable questions can limit output even when a digital step becomes cheaper. Faster manuscripts could increase volume without reliable knowledge. The model illustrates where bottlenecks might move; it does not establish that science will accelerate by a stated multiple.[1][2]
What this means for researchers—and what evidence comes next
For scientists, the pattern suggests Claude is currently most usable where an artefact can be inspected: text, code, visualisation or an analytical plan. Labs can require source-linked notes, executable code, logged model versions, independent statistical checks and a named human responsible for each output. The more autonomous the workflow, the more important it is to preserve inputs, intermediate decisions and failed attempts.
Institutions should not treat session volume as productivity. The study does not measure accepted papers, discoveries, reproducibility or social value. Leaders need outcomes reflecting their mission while protecting unpublished data, participant information and intellectual property. Access may widen inequality: fields and institutions with digitised data, compute and permissive regulation could move faster than settings where experiments, languages or local knowledge are poorly represented.
The strongest next evidence would connect preregistered workflows to time, cost, error, novelty and replication, counting checking and correction. Independent replication across providers, plans, countries, languages and disciplines would show whether the pattern is general. Unique-user denominators and longitudinal tracking could distinguish broad adoption from heavy users. Better reproducibility without hidden checking work would strengthen the case; high correction burden, privacy failures or unequal benefits would weaken it.[1][2]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring research methods: compare the finding and its evidence limits.
Can an AI choose your next experiment before you spend the compute?
A new preprint tests training intuition. We compare its narrow task with engineering, replication and architecture-search benchmarks to explain what research teams should measure next.
Read the next reportExplore the full source and evidence record
What this means for people
- Researchers can design checks around drafts, code and analysis rather than treating each session as a completed result.
- Research leaders should measure errors, replication, total checking time and access inequality—not session volume alone.
- Funders and journals need clear disclosure, data-handling and accountability rules as workflows become more autonomous.
Global context
Scientific AI adoption depends on digitisation, language resources, institutional budgets and whether regulation permits sensitive data to leave controlled environments. A session sample from one US provider cannot establish global equity. Physical facilities, data sovereignty and research culture may produce very different patterns across regions and lower-resourced institutions.
What the evidence does not yet show
- The report is authored by Anthropic, which has a commercial interest in Claude use, and is not described as peer reviewed.
- The sample covers one provider's consumer-plan sessions over two months and excludes programmatic Claude Code.
- Sessions are not unique people, projects, papers or successful outputs; no unique-researcher denominator is published.
- Classification and many SciNet characteristics are partly model-derived and cannot be independently checked from private transcripts.
- The 528 survey respondents came from 1,813 Anthropic grant applicants and likely overrepresent computational, AI-intensive researchers.
- Acceleration estimates depend on strong cost, quality and automation assumptions and are not observed productivity effects.
What to watch next
- Independent studies connecting AI use to time, error, novelty, replication and accepted outputs.
- Unique-user and longitudinal evidence showing whether adoption is broad or concentrated.
- Human-validated audits of the field and task classifiers used on private session data.
- Institutional rules for disclosure, confidential data, authorship and responsibility.
Living evidence record
Impact record IAI-1MSBSO1
Evidence stage
Announced
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
5 November 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 2 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 5 November 2026
Research question, topics and regions
How researchers use Claude across scientific fields and what session telemetry and a selected survey can establish
Topics: Scientific work · Research productivity · Product telemetry · AI agents · Research methods
Regions: Global · United States · Canada
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.