Key takeaways
- Anthropic published “AI for scientific discovery: How researchers use Claude” on November 5, 2026. This is an observational and modeling study, not a new Claude model, API, price or product-access announcement.
- The main sample contains 207,049 scientific-work sessions from August 2 through September 30: 128,914 from Claude Chat and Cowork and 78,135 from interactive Claude Code use on consumer plans.
- Anthropic reports that data analysis accounts for 33.7% of Claude sessions and 48.1% of Claude Code sessions. Writing and communication account for 32.0% of Claude sessions, while data gathering and idea generation are underrepresented relative to estimated researcher time.
- Artificial intelligence and machine learning represent 5.0% of Claude and 15.2% of Claude Code research sessions despite 1.2% of the worldwide researcher count in Anthropic’s denominator. Those are relative-use estimates, not evidence that Claude improved research quality.
- The report’s roughly 2× current-capability ceiling in some fields and 15–16× high-automation scenarios are model outputs. They depend on wide adoption, task complementarity, automation assumptions and illustrative labor and AI costs; they are not measured productivity gains or forecasts.
- Research organizations should first audit workload-shaped use, accepted outputs, verification time and missed tasks. Do not turn session share into ROI, autonomy into correctness or a modeled ceiling into a staffing plan.
Anthropic published the Claude research-use study on November 5
Anthropic Institute published “AI for scientific discovery: How researchers use Claude” on November 5, 2026. The study asks where scientists use Claude today and how that use relates to a modeled measure of AI capability. It does not announce a new model, endpoint, context limit, price or access tier.
The main evidence is a first-party classification of 207,049 sessions from August 2 through September 30, 2026: 128,914 sessions from Chat and Cowork, grouped as “Claude,” and 78,135 interactive Claude Code sessions. The appendix says the data came from Free, Pro and Max consumer plans and excludes programmatic Claude Code use. Enterprise, API and institutional deployment patterns are therefore outside this sample.
Data analysis and writing dominate current use
Anthropic reports that data analysis makes up 33.7% of Claude research sessions and 48.1% of Claude Code sessions, compared with an estimated 27.6% of researcher time. Writing and communication make up 32.0% of Claude sessions versus an estimated 16.0% of researcher time. Drafting manuscripts is the single most frequent task on both products in the report.
Data gathering is the opposite: 11.5% of Claude and 11.9% of Claude Code sessions, compared with 28.6% of estimated research time. Ideation and hypothesis generation account for 4.2% and 2.1% respectively, versus 7.1% of research time. These comparisons combine classified product use with time estimates from SciNet; they do not show that more frequent tasks are performed well or that low-use tasks should be automated.
Computational fields use Claude disproportionately
Computer science accounts for 11.9% of Claude research sessions and 32.7% of Claude Code sessions, according to the report. At the subfield level, artificial intelligence and machine learning account for 5.0% of Claude and 15.2% of Claude Code sessions while representing 1.2% of the global researcher count used by Anthropic. Mathematics, statistics and other computational subfields also show high relative use.
That pattern is consistent with the interfaces: code, text and digital data are directly actionable, while physical or interpersonal work is not. It is still a usage ratio, not a productivity result. The denominator comes from SciNet’s researcher estimates, countries without Claude availability are excluded, and low-volume subfields carry more classification and sampling uncertainty.
The capability score is a model-derived proxy
Anthropic maps each session to SciNet, a taxonomy of 34 fields, 318 subfields and 7,262 tasks. For each task, another model compares the task’s steps in pairs and a Bradley–Terry model aggregates those comparisons into a “theoretical AI capability” score. Tasks in the top capability quintile receive about twice the use predicted by researcher time on Claude and about four times on Claude Code.
The correlation is useful as a hypothesis: researchers appear to bring Claude to work that is cognitive, digital and easier to verify. But the score does not test Claude completing each task against a domain-expert gold standard. It is built from model judgments over an LLM-assisted taxonomy, and observed use can also reflect interface fit, policy, access, user selection and familiarity.
The acceleration numbers are scenarios, not forecasts
For current capabilities, the model estimates AI could take on 27–52% of time needed for a study if adopted widely, with a 38% median across fields. Under the paper’s production model, that corresponds to 1.35× in the arts, 1.41× in chemistry, 1.96× in computer science and 2.02× in statistics. Anthropic did not observe those gains in laboratories or research groups.
The larger 8× and 15–16× figures come from hypothetical scenarios where AI automates 75% or 100% of non-physical task steps, beginning with the highest capability scores. The calculation assumes an AI-agent hour costs $30, a researcher hour costs $95 and 0.2 AI hours replace each saved researcher hour. Anthropic labels these assumptions illustrative and the scenarios non-forecasts. Physical facilities, machinery, research quality, idea selection, verification capacity and field-level bottlenecks can all move the result.
The sample and classifiers set firm limits
A model first identifies scientific work, then assigns field, subfield and task labels. Anthropic says researchers retain only fixed labels rather than open-ended details, suppress cells below a minimum account threshold and weight sessions by inverse sampling probability. Those safeguards describe the analysis process; AccessAllGPT could not audit the transcripts, privacy implementation, classifier errors or account-level sampling.
The accompanying survey covers 528 US- and Canada-based applicants to Anthropic AI-for-science and related grant programs, primarily in life and health sciences. Because they sought AI research credits, Anthropic says they are likely heavier AI users than a typical scientist. The supplemental Claude Science analysis is also subject to adopter selection. Neither sample supports a population-wide claim about all researchers, products or providers.
What research leaders should do next
Start with a local activity and outcome map. For each research stage, record whether AI assists, shares or delegates work; the interface and model used; data permissions; human verification; accepted artifacts; rework; elapsed time; cost; and critical misses. Include low-use but decision-critical work such as question selection, evidence gathering and replication rather than measuring only the tasks AI already attracts.
Run a time-bounded comparison on frozen, non-sensitive work. Measure accepted analysis, traceable citations, reproducible code, reviewer time, error severity and downstream decision quality against the existing process. Ship only where evidence survives qualified review. Keep consequential scientific decisions and any physical or regulated actions behind domain-owner approval; a high session share or theoretical capability score is not an authorization boundary.
Copy-ready research AI use and outcome audit
Use this to separate observed local value from product-use counts and modeled acceleration scenarios.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Field, project stage, accountable principal investigator, users, dates, output and consequence if the work is wrong.
Ideation, reading, data gathering, analysis, design, writing and administration; current human time and bottlenecks for each.
Provider, product, model or snapshot, interface, tools, data connections, prompts, permissions and retention controls.
None, assistive, shared, delegated or autonomous; exact actions the system may take and actions reserved for people.
Representative cases, existing-process baseline, holdout set, domain reviewers, acceptance criteria and stop conditions.
Accepted analyses or artifacts, reproducibility, citation accuracy, critical errors, omissions, rework and downstream decision quality.
Elapsed time, researcher and reviewer time, model and tool cost, failures, retries, data incidents and recovery.
Important tasks with low AI use, physical or interpersonal steps, regulatory constraints and unresolved evidence.
Adopt, bounded trial, restrict or reject; eligible tasks, approval boundaries, evidence owner, expiry and re-evaluation trigger.
Primary sources
Browse the publication-wide evidence index →
- AI for scientific discovery: How researchers use ClaudeAnthropic Institute · Reviewed: November 5 publication metadata; summary; methods; field and task results; capability analysis; acceleration scenarios; discussion; footnotes; authorship and citation · Retrieved · Supports: Anthropic reports a privacy-preserving classification study of 207,049 scientific-work sessions from Claude consumer products, a supplemental Claude Science analysis and a survey of 528 researchers. Usage, capability and acceleration results are first-party findings.
- AI for Scientific Discovery technical reportAnthropic Institute · Reviewed: Full report; sample construction; task distributions; field and subfield comparisons; theoretical AI capability measure; scenarios; assumptions; limitations; footnotes · Retrieved · Supports: The downloadable report provides the exact sample counts, task shares, field comparisons, scenario assumptions and qualifications summarized in the web article.
- Appendix to AI for Scientific DiscoveryAnthropic Institute · Reviewed: Usage sample; privacy-preserving analysis; sampling weights; SciNet mapping; classifier design and validation; survey recruitment; Claude Science sample; capability measure; acceleration model and robustness analyses · Retrieved · Supports: The appendix documents that the main sample covers Free, Pro and Max consumer plans from August 2 through September 30, excludes programmatic Claude Code use, weights sessions for sampling probability and relies on model classifiers and an LLM-built task taxonomy.
- SciNet: The Anatomy of ScienceSciNet · Reviewed: Public taxonomy overview; domain, field, subfield, task and step counts; data-exploration links · Retrieved · Supports: The live SciNet site exposes the taxonomy scale used by Anthropic: six domains, 34 fields, 318 subfields, 7,262 tasks and 90,583 steps. It does not independently validate Anthropic’s Claude-session classifications or acceleration estimates.
Limitations
This is a source-backed briefing on an Anthropic-authored observational and modeling study, not an independent audit of Claude use or scientific productivity. AccessAllGPT did not inspect session transcripts, user accounts, sampling code, privacy infrastructure, classifier prompts and outputs, SciNet validation data, survey responses, Claude Science sessions, product telemetry or acceleration-model code; rerun a classifier; reproduce a usage share, correlation or scenario; measure research quality, discoveries, time saved or return on investment; compare Claude with another provider; or test any biological, laboratory or other scientific application. Consumer-plan sessions and a selected researcher survey do not represent all scientists. Session counts are not unique-user counts, classified use is not task success, model-derived capability scores are not direct task evaluations, and modeled acceleration depends heavily on adoption, automation, cost, complementarity and bottleneck assumptions.
Disclosures
AccessAllGPT received no Anthropic credentials, Claude access, session data, survey data, research credits, private artifacts, briefing, review, payment or compensation for this article. Anthropic and SciNet did not sponsor, review or endorse it. AccessAllGPT did not use, test, score or rank Claude, Claude Code or Claude Science for this report. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Anthropic or SciNet. Publication-wide relationships are listed on the disclosures page.
