- Anthropic analyzed 207,049 Claude and Claude Code science sessions from August 2–September 30, 2026, classified with a privacy-preserving taxonomy covering 7,262 tasks across 34 fields. The data describe Claude usage, not all scientific research.
- Data analysis accounts for 33.7% of Claude sessions and 48.1% of Claude Code sessions; ideation/hypothesis generation accounts for 4.2% and 2.1%, respectively. These are session classifications, not measured productivity improvements.
- Anthropic's modeled scenarios suggest some fields might double output with broad adoption and could theoretically grow much faster if non-physical work were automated. Those results are scenarios, not observed increases in discoveries, validated drugs or peer-reviewed breakthroughs.
The Signal
Research organizations want to know whether generative AI is changing the process of discovery, not simply speeding up routine clerical work. A November 5, 2026 report from Anthropic's Economics team examines the issue through a large but specific evidence source: 207,049 research sessions on Claude and Claude Code consumer plans between August 2 and September 30. The analysis assigns anonymized sessions to a structured taxonomy of 7,262 tasks spanning 34 scientific fields and 318 subfields.
The main difference across scientific fields is not just enthusiasm. Computer science, mathematics and statistics have especially high Claude use relative to their researcher populations, while medicine and physically intensive laboratory work show comparatively lower use. The paper argues that observed use tends to follow where AI can handle well-specified digital tasks, rather than where the largest future scientific benefits are imagined.
At the task level, analysis and communication dominate. Data analysis accounts for 33.7% of classified Claude science sessions and 48.1% of Claude Code sessions, while writing and communication account for 32.0% of Claude sessions. Ideation and hypothesis generation represent only 4.2% of Claude and 2.1% of Claude Code sessions. These are differences in how a particular product is used, not a measurement of the number of discoveries attributable to it.
What the evidence changes
Anthropic compares the session mix with estimates of how researchers spend their time. Data gathering accounts for 28.6% of research time in the comparison taxonomy, but only 11.5% of Claude and 11.9% of Claude Code sessions. That gap is unsurprising when tasks include collecting clinical samples, handling instruments and generating physical evidence. A language model may assist planning or interpretation, but it cannot turn a laboratory result into a measured observation without a physical workflow.
The study also surveys 528 scientist users, primarily in life and health sciences, about AI use and expectations. They rank ideation and hypothesis generation as especially important to producing significant findings, despite comparatively low use of Claude for those purposes. This contrasts what researchers consider high-value with what is currently routine and convenient. It may indicate a capability gap, user preference or a limitation in session classification.
Anthropic's specialized Claude Science interface changes the observed mix: ideation and hypothesis generation account for about 8% of its science sessions, roughly twice the general Claude share. That suggests specialized access to biological databases can support more demanding workflows. However, the specialist sample differs from general sessions, so the gap cannot be treated as proof that one interface causally doubles scientists' creativity.
The Cortist perspective
Cortist draws three separate lines: product usage, task performance and scientific outcomes. Usage can demonstrate adoption patterns. Task-level evaluations can test whether a model helps complete specific research steps. A scientifically valuable outcome requires credible evidence of a novel, replicated finding or a measurable improvement in research throughput, ideally against an appropriate baseline. The report contains the first kind of evidence and modeling relevant to the others; it does not complete the causal chain.
The authors present simple scenarios in which broad use of current capabilities could double research output in some fields. A stronger hypothetical assumption—AI carrying out all non-physical work—yields larger figures, up to fifteenfold in some computer science and statistics scenarios, compared with smaller modeled gains in bench sciences. These are conditional model outputs; quoting them as today's realized productivity would be misleading.
The analysis is both unusually broad and intrinsically selected. It observes only sessions that took place in Claude products, not researchers who never used Claude or work carried out on other platforms. It uses classifiers to infer field and task rather than direct ethnographic observation, and the companion survey draws from active users. Privacy-protecting aggregate classifications reduce risks to research confidentiality but also limit outsiders' ability to audit classification errors from raw transcripts.
- Prospective controlled studies linking AI use to verified research output, time-to-result and replication rather than session counts alone.
- Independent replication of the task classifier and examination of how multi-purpose sessions change apparent ideation rates.
- Evidence that specialized scientific agents improve novel hypothesis quality and the success rate of subsequent experiments.
- Separate measurement of physical-lab constraints, adoption costs and research quality across fields.
The largest current signal is where scientists use AI, not proof that the pace of discovery has already multiplied. Anthropic's data describe meaningful uptake of analysis and writing assistance. The more difficult accountability question is whether those tasks produce better, independently validated science.
Sources, provenance and limits
The figures above are attributed to primary reports and specified assessments. Company-run benchmark results, an external assessment with redactions and usage classifications do not provide the same kind of evidence. Cortist has not independently rerun the tests or inspected private source datasets.
Each source link identifies the party responsible for the original statement. Read the editorial standards and research methodology.
Anthropic Economics: AI for scientific discovery — How researchers use ClaudeOriginal field-and-task usage study, 5 November 2026Cortist Reality Ledger CL-0012: RNA benchmark and independent replicationRelated Cortist evidence record on reproducible scientific MLCortist: research standards and limitsEditorial methodology and source attribution policy