Anthropic's Science Study Puts AI Research Speedup at 1.35x to 2.02x

Anthropic's economics team estimated on Nov. 5 that today's AI, if scientists adopted it widely, could take over 27 to 52 percent of research time and speed science by 1.35x to 2.02x. The estimate comes from a study of 207,049 Claude sessions by Anthropic's own users, and the company labels its scenarios illustrative rather than forecasts.

The lead authors are Johannes Hermle, Ben Jones and Lukas Althoff, with six co-authors. The sessions ran from Aug. 2 to Sept. 30, 2026 on consumer plans: 128,914 from Claude chat and Cowork and 78,135 from Claude Code. Anthropic also surveyed 528 scientists, mostly in life and health sciences, recruited from applicants to its AI for Science and related grant programs.

Who is using it

Computer science made up 11.9 percent of Claude sessions and 32.7 percent of Claude Code sessions, though the field is 4.3 percent of the world's researchers, according to the report. Formal sciences use Claude more than their size predicts. Health and lab-based physical sciences use it less.

1. World's researchers4.3 %
2. Claude sessions11.9 %
3. Claude Code sessions32.7 %

Source: Anthropic, How researchers use Claude, Nov. 5, 2026; sessions classifier-labelled

The labels come from a classifier using the SciNet taxonomy of 7,262 tasks across 34 fields. Anthropic analysed only aggregate labels and reported cells with a minimum number of distinct accounts.

What the sessions are for

Usage is lopsided against the tasks that fill a researcher's week.

Task	Share of Claude sessions	Share of researcher time
Data analysis	33.7%	27.6%
Data gathering	11.5%	28.6%
Ideation and hypothesis generation	4.2%	7.1%

On Claude Code, data analysis rises to 48.1 percent of sessions. Tasks in the top 20 percent of Anthropic's AI-capability scores saw about 2.1 times expected use on Claude and about 3.9 times on Claude Code. Tasks in the bottom 20 percent saw very little.

Anthropic said ideation may be undercounted because it is woven into other tasks.

The speedup scenarios

If current capabilities were widely adopted, the report says AI could take on 27 to 52 percent of research time, with a median of 38 percent, which implies the 1.35x to 2.02x acceleration. A hypothetical full automation of non-physical work would reach about 15x in statistics and computer science and 3x to 4x in biology, chemistry and similar bench sciences.

The scenarios assume AI costs $30 an hour against $95 an hour for a researcher. They leave out field-level physical constraints, equipment costs and idea quality.

Among the 528 surveyed scientists, about 40 percent expect AI to carry out full research projects by 2030 and about 16 percent say it never will. Some 38 percent say AI has already matched top researchers at data analysis. Anthropic cautioned that respondents are likely heavier AI users than typical scientists and that Claude Science adopters are a self-selected group.

How it sits next to METR's number

The 2.02x ceiling resembles a figure from a different question. METR's Nov. 5 review of Anthropic's Risk Report gave "a range of reasonable estimates, between 1.3 and 2.8" for the speedup AI gives Anthropic's own AI research. That measures a different activity, in a different population, as of July 15.

Both rest on Anthropic's data: the usage study covers users who chose Claude, and the research METR reviewed is Anthropic's own. METR saw that material with redactions, and no party outside the two organisations has replicated either estimate. For another view of how Claude Code behaves in practice, see the report on its email-address fix and on Rivet's agent orchestrator.

What to watch

Anthropic's usage sample stops at Sept. 30 and covers consumer plans only, so enterprise and API use of the models is absent from every figure above.