Anthropic classified 207,049 researcher sessions on Claude

A research team deciding where to add Claude to a workflow now has a measured answer instead of an advert. Anthropic published a study on 5 November 2026 of how real researchers use Claude. Three co-first authors, Johannes Hermle, Ben Jones and Lukas Althoff, read 207,049 sessions on the consumer plans. The window ran from 2 August to 30 September 2026. They counted what field the session was in and what task it did.

Anthropic says it used a Claude classifier to tag each session. The tags came from a new map called SciNet. SciNet names 7,262 research tasks across 34 fields and 318 subfields. The company compared the tag counts to how real researchers spend their time, using another dataset of academic work shares.

Where Claude gets used and where it does not

Computer science made 11.9 percent of Claude sessions and 32.7 percent of Claude Code sessions. Computer scientists are 4.3 percent of the world's researchers. The AI and machine learning subfield alone made 5.0 percent of Claude sessions and 15.2 percent of Claude Code. That subfield is 1.2 percent of all researchers. Health and lab-based physical sciences showed lower take-up than the researcher count would predict.

Task shares show the same gap. Anthropic breaks four tasks into shares of sessions and shares of researcher time.

Task	Claude sessions	Claude Code sessions	Researcher time
Data analysis	33.7%	48.1%	27.6%
Writing and communication	32.0%	not reported	16.0%
Data gathering	11.5%	11.9%	28.6%
Ideation and hypothesis	4.2%	2.1%	7.1%

Data analysis and writing are over-represented. Data gathering is under-represented by a factor of 2.5. The pattern holds across fields.

The capability signal

The study then ranks tasks by how well Claude can do them on a benchmark Anthropic built. Top-20-percent tasks get two to four times the Claude use the researcher time would predict. Weaker tasks get less. Anthropic says researchers pick the tool for jobs the model already does well. The rest they do by hand. The paper says the pattern holds within fields too.

What a team should take from this

Three honest limits. First, every number is from one company's own product. Anthropic says it used privacy-preserving classifiers, but no third party has audited the method. Second, the researcher-time comparison is academic work, not what any single team does in a given week. Third, both datasets are self-selecting: Claude users are not a random sample of researchers.

Given those, the signal is still useful. A manager buying Claude for a research group should expect clear wins on data analysis and writing. They should expect small wins on hypothesis generation. They should plan separate tooling for data gathering. That is where researcher time is spent most and Claude is used least. The split lines up with a pattern seen in AI engineering work. Models do well when the input and the check are both clean. They do poorly on tasks that hunt through messy sources.

A manager can act on this today. Pick two tasks per researcher where Claude is likely to help. Measure the time saved after one month. If the number is not there, move the seat. Hand it to a team doing data analysis or writing most of the day. Those are the two slots the paper says pay off first.

Source

• AI for scientific discovery: How researchers use Claude, Anthropic