Anthropic analysis of 207,049 sessions finds Claude usage tracks AI capability, with data analysis and writing dominating scientific research

Details

Analysis of 207,049 research sessions (Aug 2–Sep 30, 2026) shows Claude usage closely correlates with theoretical AI capability scores, with tasks in the top 20% of capability scores seeing 2x usage on Claude and 4x on Claude Code relative to researcher time.
Data analysis is the most frequent task (33.7% of Claude sessions, 48.1% of Claude Code sessions), significantly overrepresented compared to its 27.6% share of researcher time, while data gathering is underrepresented (11.5% vs. 28.6% of time) due to physical barriers.
Scenario modeling suggests that if AI fully automates non-physical research steps, cognitive-intensive fields like Computer Science and Statistics could see 15–16x acceleration, whereas experimental sciences limited by physical bottlenecks would see only 3–4x.
In biomedical sciences and biology, Claude Science sessions handle more complex tasks with higher human-time equivalents (10.3 hours), higher AI autonomy (73.3%), and doctoral-level content (82.7%) compared to general Claude or Claude Code.
Survey of 528 scientists indicates that while data analysis and writing are currently most automated, ideation and hypothesis generation are ranked as the most important tasks for scientific discovery, yet remain underutilized by AI (4.2% of Claude sessions).

Key point