Anthropic Finds AI Can Take On 38% of Researcher Time, Doubling Output in Some Fields

TL;DR

Anthropic's economics team classified 207,049 research sessions from Claude and Claude Code into 34 scientific fields and 7,262 research tasks, with an LLM classifier that kept only the labels. If today's capabilities were adopted everywhere they help, AI could absorb a median 38% of the researcher time behind a study, which models out to roughly 2x output in statistics and computer science. The usage data also shows scientists spend their AI time on data analysis and writing, the categories they rank least important, while ideation, the one they rank most important, draws 4.2% of sessions.

How they counted

The sample runs August 2 to September 30, 2026: 128,914 sessions from Claude's chat interfaces and 78,135 from Claude Code, all on consumer plans, plus a supplemental batch from Claude Science. A classifier model read each transcript and assigned a field, subfield, and task from SciNet, a taxonomy of 7,262 tasks across 34 fields and 318 subfields. Only the labels were kept, and results are reported in aggregate, above a minimum number of distinct accounts.

To score where AI should in theory work well, an LLM judged pairwise comparisons of individual task steps, asking which of the two AI can do better, and the team aggregated the wins with a Bradley-Terry model, the same math behind Elo ratings. Every task landed on a percentile scale the report calls theoretical AI capability.

What the sessions actually contain

Data analysis leads everywhere: 33.7% of Claude research sessions and 48.1% of Claude Code sessions, against 27.6% of researcher time. Writing and communication is 32.0% of Claude sessions versus 16.0% of time. Data gathering runs the other way, 28.6% of researcher time but 11.5% of sessions, and ideation sits at 4.2% of Claude sessions against 7.1% of time.

A companion survey of 528 scientists, mostly in life and health sciences and all AI users, rates the same four categories by importance to producing an important result: ideation first at 35.6 points of 100, data gathering second at 23.6, data analysis 23.3, writing last at 17.5. Scientists rank ideation as the most important thing they do, then spend their AI budget on the dishes. The report offers three explanations for the gap: researchers may prefer generating ideas themselves, current models may be weaker at high-level ideation, and ideation woven into other tasks is easy for a classifier to miss.

The specialized interface moves the mix. On Claude Science, which wires the model into research databases and tools, ideation accounts for 8% of sessions, about double its Claude share and on par with its share of researcher time. Those sessions are heavier too: work worth an estimated 10.3 unassisted human-hours per session versus 4.9 on Claude and 5.9 on Claude Code, with 82.7% of use estimated to require doctoral-level training, and 73.3% of sessions run at high AI autonomy against 23.7% on Claude and 64.1% on Claude Code.

Usage follows the answer key

Tasks in the top 20% of the capability score draw about 2.1x the use their share of researcher time would predict on Claude, and about 3.9x on Claude Code. Tasks in the bottom 20%, mostly physical work, get almost none. The strongest correlate of a task's capability score is its suitability for reinforcement learning, at 0.78, followed by how cognitive the work is, at 0.69. Physical content drags it down, at −0.57, and so do interpersonal work, at −0.31, and regulatory restrictions, at −0.38.

The pattern is intuitive once you see it: AI is strongest where the homework comes with an answer key, meaning success is checkable, attempts are cheap to simulate, and feedback arrives fast. Packing a chromatography column has none of those properties, so Claude is barely used for it. The field mix shows the same skew. Computer science supplies 4.3% of the world's researchers and 32.7% of Claude Code research sessions, and the AI and machine learning subfield alone, 1.2% of researchers, supplies 15.2%.

From 38% to 15x

The scenario modeling turns usage into speedups. With current capabilities adopted everywhere they help today, AI could take on 27-52% of the researcher time behind a study, median 38%, which implies accelerations from 1.35x in the arts and 1.41x in chemistry up to 1.96x in computer science and 2.02x in statistics.

Two speculative scenarios push further. If AI automated 75% of non-physical steps, computer science reaches about 8x. If it automated every non-physical step, fields with little lab work, statistics, computer science, economics, mathematics, most of the social sciences, land around 15-16x, while bench sciences stall at 3-4x because physical steps are 38-41% of researcher time in those fields and stay human. A paper is a relay race: automating three legs changes nothing while the baton still waits on the fourth. The model prices an AI agent hour at $30 against a researcher hour at $95, which caps acceleration near 16x, and halving the cost of inference would double that ceiling. Since the fastest fields are the ones producing AI itself, the report flags the obvious feedback loop.

What researchers expect

The same 528 scientists expect a lot, fast. By 2030, 94% expect AI to match the best researchers in their field at data analysis, and 38% say it already does; 89% say writing, and 72% each for ideation and data gathering. About 40% expect AI to carry out full research projects alone by 2030, 64% by 2035, and 16% say never. The report also recalls the mid-2022 survey where the median expert put gold-level International Mathematical Olympiad performance at 2030; it happened in 2025.

What to steal

• Build where success is verifiable. The strongest adoption signal in the dataset is RL-suitability, tasks with checkable success and cheap simulated attempts. If your product's core loop can be graded automatically, it sits in the top adoption quintile. If a human must judge taste, budget for the bottom one.
• The ideation gap is a product gap, and wiring beats prompting. Claude Science doubled ideation's share of sessions without a new model by connecting the agent to the field's databases and tools. Context plumbing moved the task mix.
• Autonomy scales with tooling. Same vendor, same users: 23.7% of chat sessions run at high autonomy, 64.1% on an agentic CLI, 73.3% on the domain tool with data connections. Give an agent reliable context and users hand over more of the wheel.
• Physical and regulated work sets your ceiling. Bench-science fields spend 38-41% of researcher time on physical steps, and regulatory restriction correlates −0.38 with use. Anything that digitizes a physical step or clears a compliance barrier re-rates a whole field.
• Price the bot honestly. Even at $30 per agent hour against $95 per researcher hour, the modeled ceiling is about 16x. If your plan assumes free inference, redo the plan.

Caveats

• Anthropic measured its own customers on its own product, on consumer plans only. Enterprise and API science work, and every non-Anthropic tool, is invisible here.
• Every field and task label came from an LLM classifier, and the researcher-time benchmarks come from the same taxonomy. The appendix carries the validation; treat small gaps as soft.
• The scenarios are illustrative and the report says so itself. They assume wide adoption, skip facility limits and machinery costs, and do not model idea quality at all.
• The survey of 528 scientists skews toward life and health researchers who already use AI, so the expectation numbers run optimistic.

Key Takeaways

• Anthropic's economics team classified 207,049 Claude and Claude Code research sessions from August 2 to September 30, 2026, into 34 fields and 7,262 tasks, keeping only aggregate labels.
• Data analysis accounts for 33.7% of Claude and 48.1% of Claude Code research sessions and writing for 32.0% of Claude sessions, while ideation, ranked most important by surveyed scientists, draws 4.2%.
• Usage tracks a Bradley-Terry capability score: top-quintile tasks draw about 2.1x expected use on Claude and 3.9x on Claude Code, and RL-suitability correlates +0.78 with capability.
• With current capabilities widely adopted, AI could absorb a median 38% of researcher time per study, worth 1.35x output in the arts to 2.02x in statistics.
• Automating all non-physical work models out to about 15-16x in statistics, computer science, and economics, but 3-4x in bench sciences, where physical steps are 38-41% of researcher time.
• On Claude Science, ideation doubles to 8% of sessions and 73.3% of sessions run at high AI autonomy; 40% of surveyed researchers expect full AI research projects by 2030.

Sources: Anthropic, AI for scientific discovery: how researchers use Claude, report PDF, SciNet taxonomy, Clauding, Radar Digital