In Anthropic’s sample, analysis and writing dominate researchers’ use of Claude. The study covers scientific sessions from August 2 to September 30, 2026.
For research teams and funders, the findings offer a useful starting point for deciding where to evaluate AI assistance. They show which tasks people bring to the tool, not whether the answers are correct or discoveries arrive sooner.
Where researchers seek help
Anthropic analysed 207,049 consumer-plan research sessions: 128,914 from Chat and Cowork, grouped together in the report, and 78,135 from Claude Code. These are sessions, not a count of individual scientists or completed projects. The report’s publication date is unverified; the August–September dates describe data collection.
The company used automated classifiers to assign sessions to research tasks. Four categories show the contrast within the Chat/Cowork sample:
Research task | Share of Chat/Cowork research sessions |
|---|---|
Data analysis | 33.7% |
Writing and communication | 32.0% |
Data gathering | 11.5% |
Ideation and hypothesis generation | 4.2% |
These reported shares cover four of the study’s eight categories. Adding analysis and writing gives 65.7%, using the rounded figures—almost two-thirds of Chat/Cowork research sessions. That is a share of classified interactions, not research hours or time saved. In Claude Code, data analysis alone accounts for 48.1%.
There is another distinction behind the comparisons. The report puts data gathering at an estimated 28.6% of researchers’ time, well above its session share. But that time baseline comes from language-model estimates in the SciNet task database, which Anthropic says were checked against laboratory protocols and academic papers. It is not a time diary from the sampled users.
The pattern suggests places to investigate, but it cannot establish how much work AI completes successfully. Session length, repeat users and classification choices also complicate any attempt to turn activity counts into productivity figures.
Popular tasks and important tasks differ
In an accompanying survey of 528 scientists, respondents ranked generating ideas and hypotheses first, and gathering data second, for producing important results. The sample mainly comprised life- and health-sciences researchers who actively use AI—not a representative cross-section of all scientists.
That contrast matters: the tasks attracting the most AI interactions are not necessarily the steps researchers consider most consequential. Physical data collection can require laboratory work. Generating an idea is different, and the authors acknowledge that it may happen inside another task and escape its own classification. A small ideation share does not prove that useful ideas are absent.
The same survey reports that about 40% of respondents expect AI to conduct full research projects autonomously by 2030, while about 16% expect that never to happen. Those are respondents’ expectations, not measured probabilities or evidence that autonomous discovery has been achieved.
What the acceleration figures assume
Anthropic’s acceleration model asks how much more research output a fixed budget might support. Its scenario for widespread adoption of current capabilities reaches roughly twice the output in some fields. That is not a measured doubling of researchers’ productivity, or proof that a project would finish in half the calendar time.
The model converts reported levels of AI use into assumed fractions of researcher time taken over. For example, assistive use is assigned 25%. That conversion is a modelling choice, not a direct measurement of hours saved.
A more speculative scenario assumes AI performs every non-physical research step while humans continue physical work. Under those conditions, the model reaches roughly 15–16 times output in some fields, but only three to four times in several bench sciences. The authors explicitly describe these as scenarios, not forecasts.
A simple calculation shows how strongly the upper limit depends on costs. The report assumes an AI-agent hour costs $30, a researcher hour costs $95, and 0.2 agent-hours replace one researcher-hour. For fully substituted work, that gives 0.2 × $30 = $6, and $95 ÷ $6 ≈ 15.8.
This arithmetic helps explain the model’s roughly sixteenfold ceiling; it is not an observed return. The inputs are modelling assumptions, not a price quote for a laboratory. The model excludes machinery and experimental materials, does not capture every shared-facility constraint, and leaves the quality of research ideas unresolved.
Find the step that holds up the result
The practical implication is to separate three questions: where researchers use AI, where they expect progress, and what actually limits a particular project. The study informs the first two. A laboratory still needs evidence about the third.
Time itself has two meanings here. The SciNet documentation distinguishes hands-on researcher hours from elapsed time, including waits such as incubations or computing runs. As a practical inference, reducing analysis effort need not shorten a project whose next step depends on equipment access or physical collection.
For a team choosing an evaluation, the useful comparison is effort and elapsed time to an expert-checked result. Count prompting, review and corrections alongside software and computing costs. If assistance frees staff time but leaves the completion date unchanged, report those as separate outcomes.
The Claude-assisted vault-protein research illustrates why a structural result, an explanation of biological function and an eventual application are different milestones. This broader usage study offers a map of candidate tasks; evidence of progress still has to come from the results those tasks produce.