Anthropic Maps Claude Use in Science

Anthropic researchers analyzed 207,049 scientific research sessions on consumer plans to measure how Claude is used across research fields and tasks. The sessions ran from August 2 to September 30, 2026, and were categorized with privacy-preserving methods. The sample included 128,914 sessions from Chat and Cowork, grouped as “Claude,” and 78,135 from Claude Code. Researchers also examined a supplemental sample from Claude Science. They say only labels such as field and research task were retained; results were reported in aggregate for groups meeting a minimum account count.

Use varied by field. Computer science, engineering and medicine accounted for 11.9%, 8.4% and 7.7% of Claude sessions, respectively. Across fields, session share and global researcher share had a log-log correlation of 0.79. Formal sciences, including computer science, mathematics and statistics, had higher use relative to their researcher numbers, while health and lab-based physical sciences had lower use. On Claude Code, the correlation with researcher counts was 0.62. Computer science made up 4.3% of researchers but 32.7% of Claude Code sessions. Across 318 subfields, AI and machine learning represented 5.0% of Claude sessions and 15.2% of Claude Code sessions, despite accounting for 1.2% of researchers.

Data analysis was the most common research task, making up 33.7% of Claude sessions and 48.1% of Claude Code sessions, compared with 27.6% of researchers’ time. Writing and communication accounted for 32.0% of Claude sessions, against 16.0% of researcher time. Data gathering represented 28.6% of researcher time but 11.5% of Claude sessions and 11.9% of Claude Code sessions; tasks involving physical collection of clinical or biochemical data were especially underrepresented. Researchers used Claude less for idea generation and data gathering, although a survey ranked those tasks as most important for scientific discovery. On Claude Science, idea generation made up about twice the share of use it did on Claude.

Anthropic created a measure of theoretical AI capability by research task. Tasks in the top 20% of that measure received two to four times the Claude use predicted by the time researchers spend on them; tasks in the bottom 20%, including most physical work, received little use. The report says tasks with higher capability scores also tend to be more cognitive and better suited to reinforcement learning, where a system learns from feedback over repeated attempts.

A model in the report estimates that widespread adoption of existing AI capabilities could double research output in some fields. In a hypothetical scenario where AI performs all non-physical research work, output could rise as much as 15-fold in computer science and statistics, and three- or fourfold in bench sciences such as chemistry and biology, which face physical bottlenecks. In a survey of researchers who actively use AI, about 40% expect AI models to carry out full research projects on their own by 2030. The report also notes risks including less human oversight of dangerous or unethical research and changes to researchers’ careers and sense of purpose.

Anthropic researchers analyzed 207,049 scientific research sessions on consumer plans to measure how Claude is used across research fields and tasks. The sessions ran from August 2 to September 30, 2026, and were categorized with privacy-preserving methods. The sample included 128,914 sessions from Chat and Cowork, grouped as “Claude,” and 78,135 from Claude Code. Researchers also examined a supplemental sample from Claude Science. They say only labels such as field and research task were retained; results were reported in aggregate for groups meeting a minimum account count.

Use varied by field. Computer science, engineering and medicine accounted for 11.9%, 8.4% and 7.7% of Claude sessions, respectively. Across fields, session share and global researcher share had a log-log correlation of 0.79. Formal sciences, including computer science, mathematics and statistics, had higher use relative to their researcher numbers, while health and lab-based physical sciences had lower use. On Claude Code, the correlation with researcher counts was 0.62. Computer science made up 4.3% of researchers but 32.7% of Claude Code sessions. Across 318 subfields, AI and machine learning represented 5.0% of Claude sessions and 15.2% of Claude Code sessions, despite accounting for 1.2% of researchers.

Data analysis was the most common research task, making up 33.7% of Claude sessions and 48.1% of Claude Code sessions, compared with 27.6% of researchers’ time. Writing and communication accounted for 32.0% of Claude sessions, against 16.0% of researcher time. Data gathering represented 28.6% of researcher time but 11.5% of Claude sessions and 11.9% of Claude Code sessions; tasks involving physical collection of clinical or biochemical data were especially underrepresented. Researchers used Claude less for idea generation and data gathering, although a survey ranked those tasks as most important for scientific discovery. On Claude Science, idea generation made up about twice the share of use it did on Claude.

Anthropic created a measure of theoretical AI capability by research task. Tasks in the top 20% of that measure received two to four times the Claude use predicted by the time researchers spend on them; tasks in the bottom 20%, including most physical work, received little use. The report says tasks with higher capability scores also tend to be more cognitive and better suited to reinforcement learning, where a system learns from feedback over repeated attempts.

A model in the report estimates that widespread adoption of existing AI capabilities could double research output in some fields. In a hypothetical scenario where AI performs all non-physical research work, output could rise as much as 15-fold in computer science and statistics, and three- or fourfold in bench sciences such as chemistry and biology, which face physical bottlenecks. In a survey of researchers who actively use AI, about 40% expect AI models to carry out full research projects on their own by 2030. The report also notes risks including less human oversight of dangerous or unethical research and changes to researchers’ careers and sense of purpose.