Anthropic’s November 5 report examines 207,049 scientific sessions on Claude, Claude Code and Claude Science from August 2 to September 30, 2026, using privacy-preserving methods to classify fields and tasks without retaining the underlying research details.

Researchers used Claude most for data analysis and writing, which accounted for 33.7% and 32.0% of sessions, respectively; data analysis made up 48.1% of Claude Code sessions, while idea generation and data gathering were less common.

A survey of 528 scientists who actively use AI showed a mismatch between frequency and perceived importance: respondents used AI most for analysis and writing, but ranked idea generation and data gathering as the tasks most important to significant research results.

Claude use is concentrated in formal sciences: computer science accounts for 32.7% of Claude Code sessions despite representing just 4.3% of researchers worldwide, while health sciences and laboratory-based physical sciences use it less.

That pattern tracks AI capability: tasks in the top fifth of capability scores receive about twice the expected usage on Claude and four times the expected usage on Claude Code, while lower-scoring tasks—many involving physical work—see little use.

Capability is not the whole story, however: formal sciences, social sciences, and arts and humanities use Claude more than capability scores alone would predict, while health sciences use it less, potentially because of regulatory and integration barriers.

The report connects stronger AI performance to cognitive work and tasks suited to reinforcement learning, where systems improve through repeated attempts and verifiable feedback; physical, interpersonal and regulated work generally scores lower and sees less Claude use.

Claude Science, launched in beta on October 14, 2026, had a higher share of sessions devoted to idea generation and includes scientific-structure rendering, reproducible code-linked outputs and high-performance computing compatibility.

The specialized tool is used predominantly in biology and biomedical sciences, supporting more specialized and autonomous work; idea generation represents 8% of its sessions, roughly matching its share of researchers’ time, unlike on general-purpose interfaces.

The report presents illustrative scenarios, not forecasts: widespread adoption of current capabilities could double research output in some fields, while hypothetical automation of all non-physical research work could lift output as much as 15-fold in computer science and statistics, versus roughly three- to fourfold in bench sciences constrained by physical work.

About 40% of surveyed AI-using researchers expect AI models to conduct complete research projects independently by 2030, while the report flags risks including weaker human oversight, unethical or dangerous research, and changes to researchers’ careers and sense of purpose.