On 5 November 2026, Anthropic published the report "AI for scientific discovery: How researchers use Claude", based on an analysis of 207 049 scientific sessions in Claude and Claude Code over the period from 2 August to 30 September 2026. The study was conducted by the company's team of economists under the leadership of Johannes Hermle, Ben Jones and Lukas Althoff. The methodology relies on privacy-preserving classifiers that assign labels to sessions according to the SciNet taxonomy (34 fields, 318 subfields, 7 262 tasks), retaining only aggregated data.
The report's key finding: Claude's usage correlates strongly with the model's theoretical ability to perform a task. Tasks in the top quintile on this indicator receive 2–4 times more sessions than the share of researchers' time would predict. The bottom quintile, which includes most physical work, is barely used. This is not a coincidence but a systemic pattern: the formal sciences (computer science, mathematics, statistics) and computational subfields (AI/ML, computational biology) show relative overuse 3–7 times higher than their share among the world's researchers.
By task, the picture is even clearer. Data analysis accounts for 33,7 % of sessions in Claude and 48,1 % in Claude Code, versus 27,6 % of researchers' real time. Writing and communication — 32,0 % versus 16,0 %. At the same time, data gathering and ideation and hypothesis generation are noticeably underused (11,5 % and 4,2 % versus 28,6 % and 7,1 %). A survey of 528 active AI users confirms: it is precisely idea generation and data collection that scientists consider most important for breakthroughs, yet it is precisely here that the model still falls short of humans.
Comparison with parallel approaches from other labs reveals a divergence. OpenAI and Google DeepMind in their reports more often focus on demonstrations of specific discoveries (AlphaFold, GNoME), whereas Anthropic measures the actual distribution of workload across the entire research cycle. This makes it possible to see not "what is possible" but "what is already happening in mass use". The difference matters: demonstrations show the ceiling, while Anthropic's data show the current floor.
Scenario modeling provides quantitative estimates. With broad adoption of current capabilities, output could grow 1,35–2,02 times. In the hypothetical case of full automation of all non-physical tasks, the gain in computer science and statistics reaches 15-fold, while in chemistry and biology it is limited to 3–4-fold because of laboratory bottlenecks. These figures depend on assumptions about adoption and capability, but they point to a potential feedback loop: accelerating AI research itself accelerates the development of more powerful models.
The methodology has limitations that the authors explicitly acknowledge. The sample is limited to consumer plans; users of Claude Science (a specialized tool) show a different structure — a higher share of ideation and autonomy. The classifiers are trained on the SciNet taxonomy and may miss subtle or interdisciplinary tasks. There is no independent third-party verification, although the privacy-preserving approach makes it possible to reproduce the aggregated metrics given access to similar data.
Open questions concern long-term effects. Will the automation of non-physical tasks increase pressure on physical infrastructure or, conversely, allow researchers to spend more time on experiments? How will the career path of young scientists change if routine tasks go to models? And most importantly — will human understanding and responsibility for results be preserved when the share of AI-generated content in papers grows many times over.
Anthropic's report shows that the current wave of adoption is already redistributing researchers' efforts toward tasks where models are strongest, and thereby setting a trajectory of acceleration that will be uneven across disciplines.
