Skip to content

Anthropic

Scientists use Claude least where they say it matters most

8 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Claude Code versions and changes →
Image: anthropic.com

Anthropic on November 5 published an analysis of 207,049 research sessions on Claude and Claude Code. It finds that scientists use Claude most for data analysis and writing, and least for ideation and data gathering, the two tasks surveyed researchers rank as most important for discovery. The usage shares are measured from real sessions. The claims that today’s AI could double output in some fields, and that automating all non-physical work could lift computer science and statistics as much as 15-fold, come from an illustrative model. When we rebuilt that model, its ceiling turned out to depend on an AI runtime figure that the authors say no data source measures.

The sessions in Anthropic’s report come from consumer plans between August 2 and September 30. Of these, 128,914 came from Chat and Cowork, which the report groups together as “Claude,” and 78,135 came from Claude Code. A language model read each transcript and kept only two labels, field and task, drawn from a taxonomy of 7,262 research tasks. A survey of 528 scientists runs alongside the session data. Most respondents work in the life and health sciences and already use AI. They were recruited from applicants to Anthropic’s AI for Science grant programs and work at institutions in the US and Canada.

Where the sessions go

Data analysis takes 33.7% of Claude sessions and 48.1% of Claude Code sessions. By the report’s estimate, it takes 27.6% of researchers’ working time. Writing and communication take 32.0% of sessions on Claude against 16.0% of researcher time. Drafting manuscripts is the most frequent single task on both products: 18.6% of Claude sessions and 11.1% on Claude Code. On Claude, reviewing literature (6.8%) and responding to peer reviewers (6.7%) come next.

Data gathering accounts for 28.6% of researcher time but only 11.5% of Claude sessions and 11.9% on Claude Code. Ideation and hypothesis generation get 4.2% on Claude and 2.1% on Claude Code, against 7.1% of researcher time. The authors give different reasons for the two gaps. Much of data gathering is physical, and purely digital collection, such as pulling data from the web, can even be overrepresented. For ideation they offer several explanations without choosing one. Scientists may prefer to come up with ideas themselves. The classifier may miss ideation that is woven into other tasks. Or AI may be genuinely weaker at choosing a research question and designing a study to answer it.

Computer science makes up 4.3% of the world’s researchers and 32.7% of Claude Code sessions. Sessions are assigned to the field of the research being done, whatever the tool, so coding for other computational sciences does not count toward computer science. Artificial intelligence and machine learning alone, with 1.2% of researchers, takes 15.2% of Claude Code sessions and 5.0% of Claude sessions. Medicine is the third-largest field on Claude, with 7.7% of sessions, but it still uses Claude less than its number of researchers would predict.

To score how well AI can handle each task, Anthropic had a language model compare pairs of task steps and judge which one AI does better, then combined those comparisons into one score per task. Tasks in the top fifth draw about twice the use their share of researcher time would predict on Claude, and about four times on Claude Code. The bottom fifth, which includes most physical work such as purifying reaction mixtures, barely shows up. The score has a 0.78 correlation with a separate index of how well a task suits reinforcement learning, where a model gets better by trying a task again and again and being rewarded when it succeeds. That kind of training works best when success is easy to check and feedback comes quickly. Language models produced both the score and the index.

What Claude Science changes

On Claude Science, ideation makes up 8% of sessions, roughly in line with its share of researcher time. Claude Science is Anthropic’s research app that connects Claude to scientific databases and tools, and over half its sessions are in biomedical sciences and biology. In those two fields, one task, “propose mechanistic models for protein–protein or protein–ligand interactions,” is 5.6 times more likely to appear than on Claude. Over 70% of sessions run with a high degree of AI autonomy, compared with 23.7% on Claude and 64.1% on Claude Code. Over 80% of the work is estimated to need doctoral-level training, against about 50–70% on the other interfaces.

The authors add a caveat to the comparison. Any paid user can open Claude Science, but the people who adopt it are likely those with more specialized or demanding work, so the report describes the differences as patterns of use. The app is in beta for macOS, Windows and Linux on Pro, Max, Ultra, Team and Enterprise plans. Anthropic’s FAQ says it runs on the same Claude models a plan already includes. Team is one of those plans, and since August, principal investigators at academic and nonprofit institutions can put their group on a free Claude Team plan for scientists, which opened with 10,000 seats.

How far do 2x and 15x go?

Scenario 1 assumes current capabilities are widely adopted and applied to every step where they help today. AI then takes on 27–52% of the researcher time a new study needs, with a median of 38% across fields. Output rises by 1.35x in the arts and 1.41x in chemistry, up to 1.96x in computer science and 2.02x in statistics. Even the most AI-ready fields reach only about 2x, because today’s use is mostly assistive or shared and researchers still spend time on nearly every step. Lab sciences such as chemistry, biomedical sciences and materials science spend 38–41% of researcher time on physical steps, against 4% in the median field.

In the third, hypothetical scenario, AI fully automates every non-physical step. Fields with little physical work reach about 15–16x. These include statistics, computer science, economics, mathematics, psychology, and most of the social sciences and humanities. Biology, chemistry, materials science and other lab-heavy fields reach 3–4x, held back by physical steps that take about 40% of researcher time. The authors call these “illustrative scenarios, not forecasts.” The ceiling near 16x comes from cost. They assume an hour of AI agent time costs $30 and an hour of a researcher’s time $95, and that AI runs 0.2 hours for each researcher hour it saves. A fully automated task therefore still costs about a sixteenth of the human version.

About 40% of the surveyed researchers expect AI to carry out full research projects on its own by 2030, and about 64% by 2035; 16% say it will never happen. On data analysis, 38% say AI already matches the best researchers in their field.

Our analysis

What we checked

On November 7 we rebuilt the report’s acceleration model from the calibration in its Appendix and recomputed the published figures. We also checked the Claude Science details against Anthropic’s current product pages.

How we check stories

The published accelerations match the model as we rebuilt it. The 15-fold ceiling, though, depends just as much on how long AI runs for each researcher hour it saves as on the $30 price, and the Appendix says that runtime is an assumption with no data behind it. The report also says halving the cost of AI would double the rate of progress, which holds only where physical work is close to zero.

  • Our version of the model gives a ceiling of 15.83x, in line with the report’s “about 16x.” For fields with 38–41% physical steps it gives 3.2–4.2x, against the published 3–4x. For Scenario 1 it gives 1.34x for the arts and 1.95x for statistics, against the published 1.35x and 2.02x. Our Scenario 1 figures are lower bounds, so published figures slightly above them fit the model. The equation appears only as an image in the PDF. We rebuilt it, and the exact match on the ceiling is how we checked the rebuild.
  • The Appendix calls the $30 “an illustrative value based on public API list prices for Claude Opus 5.5” and says it “is not a measured average cost of Claude use or of Anthropic’s costs.”
AI price per hourAI hours per researcher hour savedCeiling
$300.2 (report)15.8x
$300.56.3x
$301.03.2x
$600.27.9x
$150.231.7x

Take a field where 40% of the work is physical steps that AI does not speed up. Halving the AI price from $30 to $15 raises its Scenario 3 acceleration from 3.30x to 3.90x, about 18%, and doubling the price to $60 lowers it to 2.66x. That is our inference from the rebuilt model.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.