Anthropic Analyzes 207,000 Sessions of Claude Use in Scientific Research

Anthropic analyzed 207,000 sessions involving Claude and Claude Code in scientific fields. According to the report, use relative to the number of researchers was relatively high in the formal sciences, including computer science, mathematics and statistics, and lower in medicine and the physical sciences involving experiments. Data analysis and writing were the main uses.

Anthropic's analysis found that data analysis and writing accounted for much of Claude's use, while use for idea generation and data collection—tasks researchers consider important—was limited. Evaluations of AI for scientific research need to distinguish usage volume from whether the technology supports the tasks researchers value.

The report shows which tasks researchers actually use Claude for, not how much Claude has accelerated scientific discovery. In the analysis of 207,000 sessions, data analysis accounted for about one-third of Claude use and nearly half of Claude Code use, while use for idea generation and data collection, which researchers rated as important, was limited. Usage frequency therefore cannot be directly equated with research value or outcomes. The analysis observed use for specific tasks; this information does not show whether research quality improved or the time to discovery was reduced. Differences across fields also do not in themselves indicate differences in capability. The finding that use relative to the number of researchers was higher in the formal sciences and lower in medicine and the experimental physical sciences shows that Claude use is not equally widespread across fields. However, the information provided does not explain why usage differs by field or how access to experimental facilities or data affects use. These factors cannot be asserted as causes of the differences. The approach to measuring actual usage differs from the measure of “how much of a job’s tasks a robot can perform” addressed in Anthropic Publishes a Measure of How Much of a Job Robots Can Take On. The former concerns how researchers use AI, while the latter concerns the range of tasks robots can perform. Both offer a basis for evaluating concrete tasks rather than abstract claims about capability. However, their subjects and methods differ, so they cannot be directly compared. Future evaluations will need to consider not only session counts and task categories, but also how closely use for idea generation and data collection aligns with the challenges researchers prioritize, and how to measure research outcomes after use. The summary provided does not specify the period covered, how sessions were categorized, the exact usage rates by field or how usage corresponded with research outcomes. Until those details are available, differences in usage should be read within the bounds of what was observed, rather than used to infer effects on scientific productivity.