Anthropic has published the clearest picture yet of how scientists actually use its AI, and the pattern is not the one the hype suggests. Across 207,049 research sessions on Claude and Claude Code between August 2 and September 30, 2026, data analysis dominated, accounting for about a third of Claude sessions and nearly half of Claude Code sessions. Writing came next, while the work researchers themselves rank as most important — generating ideas and gathering data — drew the least use. Anthropic’s scenarios suggest that, widely adopted, today’s AI could roughly double research output in the most AI-ready fields.
What scientists actually do with Claude
Data analysis is where Claude earns its keep. It accounts for 33.7% of sessions on Claude and 48.1% on Claude Code, against 27.6% of the time researchers spend on the task. Writing and communication run a similar pattern, taking 32.0% of Claude sessions against 16.0% of researcher time. Use follows what works today.
The gaps sit exactly where the science matters most. Data gathering takes 28.6% of researcher time but only 11.5% of Claude sessions and 11.9% of Claude Code sessions, largely because so much of it happens at a bench or in the field.
Ideation and hypothesis generation fare worse still, at 4.2% of Claude sessions and 2.1% on Claude Code, against 7.1% of researcher time. Yet the 528 researchers Anthropic surveyed ranked idea generation first and data gathering second for producing important results, a mismatch the report puts at the centre of its findings.
The fields that lean in hardest
Adoption is lopsided by field as well as by task. Computer science supplies 11.9% of Claude sessions and 32.7% of Claude Code sessions, despite making up only 4.3% of the world’s researchers. Engineering and medicine follow on Claude at 8.4% and 7.7%. The AI and machine learning subfield alone takes 15.2% of Claude Code sessions from 1.2% of researchers.
The pattern is consistent: the more a field’s work lives in text and code, the more Claude gets used. Formal sciences such as mathematics and statistics sit at the top, while lab-heavy fields such as chemistry sit lower, held back by physical work and medical-data rules. Claude Code concentrates the skew, since its work is almost entirely computational.
Use follows capability
Anthropic tested whether usage tracks what AI can actually do, and the answer is yes, closely. It scored each of 7,262 research tasks in its SciNet taxonomy for theoretical AI capability, then compared use against the time researchers spend on each task.
Tasks in the top fifth of capability scores draw about twice the expected use on Claude and four times on Claude Code, while tasks at the bottom are barely used at all. The link is strongest where reinforcement learning applies easily: capability scores correlate 0.78 with how suitable a task is for learning from repeated, verifiable attempts.
The skew shows up inside Anthropic itself. Claude now writes 90% of the company’s own code, according to a filing, which helps explain why coding dominates Claude Code’s research sessions. The tool is used hardest exactly where its users already live: in repositories, tests and scripts, rather than in the laboratory.
Claude Science goes deeper
The picture changes with Claude Science, Anthropic’s specialised tool for researchers, launched in beta on October 14. Idea generation makes up about 8% of its sessions, roughly twice the share on Claude and close to the share of researcher time the task takes.
One specialised job, proposing mechanistic models for how proteins bind, is 5.6 times more likely on Claude Science than on Claude. The tool connects to research databases, combining protein structures and trial data into a single workflow, with longer, more expert sessions than the general tools.
That specialisation is already showing in Anthropic’s own biology work. Claude helped crack a 40-year protein mystery in four months, identifying the first vault-like particle structure in bacteria, a result the company published this month. It is the kind of deep, database-heavy task the study finds Claude Science best suited to, and the hardest for a general chatbot.
How much faster could science go?
Anthropic’s scenarios put numbers on the prize, and they come with a warning that they are models, not forecasts. With today’s capabilities widely adopted, AI could take on 27% to 52% of the researcher time needed for a new study, a median of 38% across fields.
That implies research output rising about 1.96 times in computer science and 2.02 times in statistics, but only 1.35 times in the arts. The gap is physical work: bench sciences spend 38% to 41% of researcher time on steps no chatbot can perform, and that share alone explains about a third of the difference between fields.
Greater automation widens the split further. If AI handled 75% of non-physical steps, output could rise about eightfold in computer science and sevenfold in statistics, against two to three times in bench sciences. Full automation lifts cognitive fields to roughly fifteenfold, while bench sciences stay near three to fourfold.
What researchers expect
Researchers themselves expect more, and soon. Over half of those surveyed expect AI to match the best researchers at data analysis by 2027, and 38% say it already has. By 2030, 94% expect it to match the best at data analysis, 89% at writing, and 72% each at ideation and data gathering.
About 40% expect AI to carry out full research projects alone by 2030, and 64% by 2035, though 16% say it will never happen. Expectations for autonomous breakthroughs run more cautiously still, with about half of researchers expecting those only after 2035, if ever.
Our Read
The study’s most useful finding is its mismatch. Scientists use AI most where it already works and least where it would matter most, and the tools built to close that gap — specialised, database-connected assistants — are only now arriving.
Anthropic’s own numbers argue that the next leap in AI for science will not come from a bigger general model, but from tools shaped around specific fields and their hardest tasks. Rival labs measure the same frontier differently: Google’s Gemini currently leads a science-image benchmark, while independent tests keep finding AI capability improving.
Why It Matters
This matters because research is where AI’s promises are largest and its evidence thinnest. A study built on 207,000 real sessions replaces anecdotes with a map of what the technology actually does in the lab, and the map says the gains so far are real but narrow.
The fields that could accelerate most are the ones that build AI itself, which would feed faster AI research back into faster models, a loop Anthropic flags as both the opportunity and the risk. The same loop is why the report’s warnings about oversight and dangerous knowledge deserve as much attention as its speed estimates.
Background
Anthropic analysed the sessions with privacy-preserving classifiers that label each transcript by field and task without retaining its contents, using a taxonomy of 34 fields, 318 subfields and 7,262 tasks. The study is the company’s most detailed public account of how its tools are used in science, published as its listing approaches and its science programme widens.
Source: Anthropic’s report
0 Comments