Economic Model Puts Statistical Research Output at 2.02 Times Baseline

An economic model in Anthropic’s analysis estimates statistical research output could reach 2.02 times its no-AI baseline at the same budget under broad adoption of current AI capabilities, highlighting uneven potential gains across scientific fields.

In the same-budget scenario, output reaches 1.96 times baseline in computer science, 1.41 times in chemistry and 1.35 times in the arts. The figures describe potential output, not productivity gains already measured in laboratories or universities.

Anthropic published the analysis Nov. 5. It covered 207,049 research-related conversations on consumer plans from Aug. 2 to Sept. 30, 2026: 128,914 on Claude and 78,135 on Claude Code. A model tagged each conversation by research field and task, and the research team received those labels without seeing the conversation content.

Data analysis made up 33.7% of Claude conversations and 48.1% of Claude Code conversations, compared with 27.6% of researchers’ time. Writing and communication accounted for 32% of Claude conversations, versus 16% of researchers’ time. Data gathering represented 11.5% of Claude conversations, compared with 28.6% of researchers’ time. Idea generation and hypothesis work accounted for 4.2% of Claude conversations and 2.1% of Claude Code conversations, against 7.1% of researchers’ time.

In a separate survey, 528 scientists who actively use AI — most in life and health sciences — ranked idea generation and data gathering as the two tasks most important to producing significant research results.

In the current-capabilities scenario, AI could handle 27% to 52% of the researcher time needed for a new study, with a median of 38% across fields. Other scenarios assume AI takes over 75% or all nonphysical research steps; Anthropic described them as illustrations, not predictions. Under the 75% scenario, modeled output reaches about eight times baseline in computer science, seven times in statistics and six times in economics. Experimental and medical fields reach two to three times baseline.

With all nonphysical work automated, modeled output rises to 15 to 16 times baseline in statistics, computer science, economics, math, psychology and most humanities and social sciences. Biology, chemistry, materials science, agriculture and biomedical sciences reach three to four times baseline. Researchers in several of those lab and field areas spend 38% to 41% of their time on physical tasks, compared with a 4% median across fields.

The model assumes an AI agent costs $30 per hour and a researcher $95 per hour, helping cap modeled output at about 16 times baseline even under full automation. It excludes equipment and experimental-material costs.