Scientists use Claude least where they say it matters most

Anthropic’s report, published November 6, studies 207,049 research sessions on Claude and Claude Code. It finds that scientists mostly use the tools to analyze data and write, and use them least for generating ideas and gathering data, the two tasks a companion survey of scientists ranked most important for discovery. The usage shares are measured from real sessions, while the speedups likely to get quoted (roughly double output in some fields with today’s tools, up to 15-fold if AI took over all non-physical work) come from scenarios the authors call illustrative.

The sessions come from consumer plans between August 2 and September 30, 2026: 128,914 from Chat and Cowork, which the report groups as “Claude”, and 78,135 from Claude Code. A classifier model labeled each transcript with a field, subfield and task from SciNet, a taxonomy of 7,262 research tasks across 34 fields and 318 subfields. Those labels are the only thing the team kept, and results are reported only for cells with a minimum number of distinct accounts.

Where the sessions go

Data analysis takes 33.7% of sessions on Claude and 48.1% on Claude Code, while researchers spend 27.6% of their time on it. Writing and communication is 32.0% of Claude sessions against 16.0% of researcher time. Data gathering runs the other way: 28.6% of researcher time, but only 11.5% of sessions on Claude and 11.9% on Claude Code. The gap is widest for physical work such as collecting clinical or biochemical data, and collecting online data can even be overrepresented. Ideation and hypothesis generation gets 4.2% on Claude and 2.1% on Claude Code, against 7.1% of researcher time. The time shares are SciNet estimates. An LLM classifier produced them, and they were validated against wet-lab protocols that log time per task and against how often tasks appear across thousands of papers.

The authors don’t pin the ideation gap on one cause. Scientists may prefer to come up with ideas themselves or may see AI as limited there. The data may also undercount ideation, because it is woven into other tasks and hard to classify. And AI may simply be weaker at high-level ideation, such as choosing which question to pursue.

By field, computer science makes up 4.3% of the world’s researchers but 32.7% of Claude Code sessions. The artificial intelligence and machine learning subfield is 1.2% of researchers and 15.2% of Claude Code sessions (5.0% on Claude). Medicine is the third-largest field on Claude by volume, at 7.7% of sessions, yet its usage is still low relative to its size; the report notes that its researchers spend much of their time with patients.

The Claude Science numbers

Ideation looks different on Claude Science, where it accounts for 8% of sessions, roughly on par with the share of researcher time it takes. That sample is supplemental, and biomedical sciences and biology together make up over 50% of it. The task “Propose mechanistic models for protein–protein or protein–ligand interactions” is 5.6 times more likely to appear there than on Claude. In those two fields, Claude Science sessions hold work that would take nearly twice as many human hours without AI as work on Claude and Claude Code, and they run more autonomously. The authors call this “consistent with specialized AI tools being used for more demanding tasks.”

MIT Technology Review described Claude Science as a product meant to support scientific research the way Claude Code supports software engineering. It launched in June. Global Healthcare Magazine says it is an interface on existing models such as Claude Opus 4.8, where a coordinating agent hands work to domain specialist agents and a reviewer agent checks citations, calculations and outputs, and that it launched in beta for Pro, Max, Team and Enterprise users.

Where does the 15x come from?

The capability measure is itself a model’s judgement. An LLM classifier was shown pairs of task steps and asked which one AI performs better, and the answers were combined with a Bradley–Terry model, the method behind chess Elo ratings. Tasks in the top 20% on this score get about twice the use their researcher time predicts on Claude and about four times on Claude Code, while the bottom 20%, mostly physical work, barely gets used. The score correlates at 0.78 with a separately classified index of how well a task suits reinforcement learning.

The first scenario assumes today’s capabilities are widely adopted, treating the surveyed researchers’ reported AI use, as heavy adopters, as a proxy for what AI can do now. On that basis AI could take on 27–52% of the researcher time behind a new study, with a median of 38% across fields, for accelerations from 1.35x in the arts to 2.02x in statistics. Lab and field sciences spend 38–41% of researcher time on physical steps, compared with 4% in the median field, and that alone explains about a third of the variation in AI’s share.

The 15-fold figure comes from the third scenario, a hypothetical where AI automates every non-physical step. Statistics, computer science, economics, mathematics, psychology and most of the social sciences and humanities reach about 15–16x. Biology, chemistry, materials science, agricultural sciences and biomedical sciences reach only 3–4x, because physical steps stay at about 40% of researcher time unless labs are automated. With AI automating 75% of non-physical steps instead, computer science reaches about 8x.

Cost sets the ceiling. The model prices an hour of AI agent time at $30, today’s public list price, against $95 for a researcher hour, and assumes AI runs 0.2 hours for every researcher hour it saves. Under those assumptions a fully automated task still costs about one-sixteenth of what a researcher would, so accelerations approach about 16x and stop there; halving the cost of AI would double the rate of progress, the authors write. They also leave out field-level limits such as time on facilities like the Large Hadron Collider, the cost of machinery and experimental inputs, simulations that could replace experiments, and the quality of ideas. “These are illustrative scenarios, not forecasts,” they write.

What the surveyed researchers expect

The survey covers 528 scientists, mostly in the life and health sciences and all active AI adopters. Already 38% say AI matches the best researchers in their field at data analysis. By 2030, 94% expect it to match them at data analysis, 89% at writing and 72% at both ideation and data gathering. About 40% expect AI to carry out full research projects on its own by 2030 and about 64% by 2035, while 16% say it will never happen. Roughly half expect AI to produce breakthroughs autonomously only after 2035, if ever.

The authors set those answers against an earlier forecast. In a mid-2022 survey, domain experts put the median date for AI reaching gold-medal level at the International Mathematical Olympiad at 2030, and superforecasters put it at 2035. AI got there in 2025. The exceptions to that pattern of underestimating AI, they add, have come mainly in work that has to happen in the physical world, such as AI-assisted lab tasks in biology.