Skip to main content
Back to timeline
AnthropicSource publication:

Anthropic tallies 207,000 research sessions: data analysis and writing dominate, while ideation and data gathering lag

Synopsis

Anthropic used privacy-preserving methods to label 207,049 Claude and Claude Code research sessions from August to September 2026 by field and task under the SciNet taxonomy. Data analysis accounted for 33.7% of Claude sessions and 48.1% of Claude Code sessions, while ideation and hypothesis generation, which researchers rank as most important, took 4.2% and 2.1%. A separate theoretical AI capability score showed tasks in the top 20% drew two to four times the Claude use their share of researcher time would predict.

Source-provided article image: How scientists use Claude \ Anthropic
Anthropic

Interpretation

Where researchers use Claude does not simply follow the size of each field's researcher population; it skews toward computationally intensive work done through text and code. By session share, the three largest fields on Claude are computer science at 11.9%, engineering at 8.4% and medicine at 7.7%; relative to global researcher counts, the formal sciences use it more, while lab-based physical sciences such as chemistry and medicine use it less. The log-log correlation between session share and researcher share is 0.79 for Claude and 0.62 for Claude Code; computer science makes up only 4.3% of the world's researchers but 32.7% of Claude Code sessions.

By task, Claude is heavily used for data analysis and writing, while ideation and data gathering, which researchers themselves rank as most important, are used less. SciNet estimates researchers spend 27.6% of their time on data analysis, 16.0% on writing and communication, 28.6% on data gathering and 7.1% on ideation and hypothesis generation. Data analysis accounts for 33.7% of Claude sessions and 48.1% of Claude Code sessions; writing and communication for 32.0% of Claude sessions; data gathering for only 11.5% and 11.9%; ideation and hypothesis generation for 4.2% and 2.1%.

Perspective

The analysis covers 207,049 consumer-plan Claude and Claude Code sessions plus a supplemental Claude Science sample, with a companion survey of 528 researchers mainly in the life and health sciences in the US and Canada who applied for AI research credits. The acceleration scenarios are illustrative, not forecasts, and assume AI agent time costs $30 per hour against $95 for a researcher hour, with 0.2 hours of AI run per researcher hour saved.

The authors note that the acceleration calculations omit facility-level limits on experiments, the cost of machinery and experimental inputs, simulations that could replace physical experiments, and the quality of ideas; better AI hypotheses would raise acceleration, while uniquely human research taste would lower it. The survey sample is heavy AI users, treated as a proxy for what AI can do today. Ideation use may be undercounted because it is woven into other tasks and hard to classify.