On 5 November Anthropic’s economics team published “AI for scientific discovery: How researchers use Claude,” a report with a 45-page appendix. The team classified 207,049 Claude and Claude Code sessions from consumer plans, dated 2 August to 30 September 2026, by scientific field and research task. The line that got quoted was the forward-looking one: “If AI could carry out all non-physical research work, output in fields such as computer science and statistics could rise as much as 15-fold, while output in bench sciences such as chemistry and biology, which are held back by physical bottlenecks, would rise three- or fourfold.”
I read the report and the appendix. The usage numbers are the most useful public data I’ve seen on what scientists actually do with a model. Two of the report’s bigger claims rest on less than they appear to. The finding that “scientists appear to use AI where it works best” sets a model’s labels against another model’s judgments. And the 15-fold figure mostly restates a cost assumption the authors chose and say they chose.
What the sessions show
The sampling is straightforward. A screener flags sessions that involve scientific work, and a second classifier assigns each one to a field, subfield and task in SciNet, a taxonomy with 7,262 tasks across 34 fields and 318 subfields. The main sample is 128,914 sessions on Claude (chat and Cowork) and 78,135 on Claude Code. Only the labels leave the privacy pipeline.
The field split is lopsided in the expected direction. Computer science is 4.3% of the world’s researchers and 32.7% of Claude Code sessions. Artificial intelligence and machine learning, a single subfield with 1.2% of researchers, accounts for 5.0% of Claude sessions and 15.2% of Claude Code sessions. Medicine is the third-largest field by Claude sessions at 7.7%, but it is below the line once you adjust for how many medical researchers there are.
The task split is more interesting. The report compares each category’s share of sessions with SciNet’s estimate of how much researcher time that category takes.
Writing gets twice its share of researcher time. Data gathering gets well under half. On Claude Code the skew is sharper: data analysis is 48.1% of sessions, and ideation is 2.1%. A footnote gives the single most common task on both products: drafting manuscripts, at 18.6% of Claude sessions and 11.1% of Claude Code sessions.
This matches outside evidence. Liang and colleagues, in Nature Human Behaviour, estimated the share of LLM-modified text in 1,121,912 papers and preprints from January 2020 to September 2024 and found the fastest growth in computer science, up to 22%, against up to 9% in mathematics and the Nature portfolio. A writing-heavy, computing-heavy pattern shows up whether you look at the papers or at the sessions.
The report’s own survey adds the uncomfortable part. Asked which kinds of work matter most for an important result, respondents ranked ideation and hypothesis generation first and data gathering second. Those are the two categories with the least use. The report offers several reasons, including that ideation is “woven into many other research tasks and is difficult to classify,” which means the classifier may miss it.
Use tracks capability, as judged by a model
The report’s central explanatory claim is that use follows capability. “The tasks AI is best suited to (the top 20% on this measure) get two to four times as much Claude use as the time researchers spend on them would predict.” It’s an intuitive result, and the evidence for it is less independent than the sentence suggests, because every quantity in it is produced by a language model.
The task labels come from a Sonnet 5 classifier. With no ground truth for real sessions, the team tested it on 5,000 synthetic transcripts generated with known labels. It got 83% of fields right, 74% of subfields and 69% of tasks. The appendix adds that “about half of the transcripts still echo some of the label’s wording after the rewrite,” and that excluding those lowers task accuracy to 64%. Random assignment would score 0.5% on tasks, so this is a real signal. It also means roughly one session in three may be counted under the wrong task.
The time estimates come from SciNet, built by Cristian Alamos and Lukas Althoff, one of this report’s lead authors. Researcher hours per task are “LLM-based time estimates validated via (1) time recorded on research protocols for wet-lab sciences from protocols.io and (2) the number of papers that contain a task.” Those checks are reasonable, but protocols.io records bench steps, which are the steps least relevant to how anyone uses Claude.
The capability score is the most interesting construction. A judge model is shown two research steps and picks the one for which “AI’s typical output without human help would come closer to what an experienced researcher in that field would accept without substantial correction.” Over 317,560 such comparisons, a Bradley-Terry model ranks every step. The appendix says the judge can also declare that AI cannot do a step, and 16% of steps get that answer in both of their first two comparisons. The appendix doesn’t say which model was the judge.
None of this makes the correlation wrong. It does change what it means. Scientists choose to bring Claude the tasks that they, and the culture around them, believe Claude handles well. A judge model trained on that same culture’s text is asked which tasks AI handles well. Agreement between the two is evidence that a model’s picture of AI’s strengths matches users’ picture of AI’s strengths. Whether measured performance matches use is a separate question, and the method doesn’t reach it. The report’s second piece of supporting evidence, that a task’s capability score correlates 0.78 with a separately classified RL-suitability index, is a third model judgment lining up with a second.
I made a similar point about a tax benchmark where Claude models also did the grading. There the issue was family bias in the scores. Here it’s quieter: the judge and the users may share the same prior, and the method can’t separate a shared prior from a true capability. The report’s own comparison is with Eloundou and colleagues’ “GPTs are GPTs” exposure measure in Science, where GPT-4 rated O*NET tasks. The appendix says the new score correlates with that rubric but ranks tasks more finely, and that 58% of SciNet tasks would be “directly exposed” under the old rubric with no further distinction. Finer ranking is a gain, but it’s still a model’s ranking.
A check against task outcomes would settle much of this. Session success rates, or a sample of tasks scored by people in the field, could be lined up against the capability percentiles. The report doesn’t have that, and given its privacy design it may not be able to.
Who answered the survey
Scenario 1, the “current capabilities” estimate, is built from the survey, not from sessions. The appendix is clear about who answered. The team “contacted 1,813 researchers by email” in September 2026, and “a total of 528 completed the survey,” a 29% response rate. All were applicants to Anthropic’s token grant programs, “drawn primarily from biological and medical sciences,” leaning “toward computational research,” and “expected to be stronger users of AI models in research, given their selection into the grant program.”
The report says it treats their reported use “as a proxy for what AI can do today,” and it says why: “Because respondents are likely to be heavy AI adopters.” The survey’s distributions are then spread across every field. In ideation, the lowest-rated 17% of steps get no AI share and the next 53% get 25%, and so on up the scale, from “none” to “autonomous.” Those five levels are mapped to 0%, 25%, 50%, 75% and 100% of researcher time saved. Nothing in the survey measures that mapping; the team chose it.
The result is an AI share of 27% to 52% of researcher time by field, a median of 38%, and accelerations from 1.35x in the arts to 2.02x in statistics. That’s what heavy adopters from biomedicine report, mapped onto historians and statisticians by the judge model’s rankings. It may well be roughly right. It is a scenario, and the report calls it one.
Where fifteen comes from
The bigger numbers come from Scenarios 2 and 3, in which AI fully automates 75% or 100% of non-physical steps. The model behind them is an ideas production function from Benjamin Jones, one of the lead authors: a paper needs many complementary tasks, and making one cheap shifts effort to the others. That idea has a long history in the literature on research productivity. Jones’s 2009 paper on the “burden of knowledge” in the Review of Economic Studies argued that rising specialization and teamwork are how innovators cope with a growing stock of knowledge, and Wuchty, Jones and Uzzi showed in Science in 2007, over 19.9 million papers, that teams “increasingly dominate solo authors in the production of knowledge.” Bottlenecks between complementary tasks are the right frame for asking how much any single tool can speed things up.
The ceiling, though, is set by three numbers in the appendix’s calibration section. A researcher-hour costs w = $95, which is “the median salary in the 2023 Survey of Doctorate Recipients ($155,000 in 2026 dollars) plus 28% benefits, over 2,080 hours.” An hour of an agent costs μ = $30, “an illustrative value based on public API list prices for Claude Opus 5.5. It is not a measured average cost of Claude use or of Anthropic’s costs.” And because “no data source measures AI running time by step,” the team assumes AI runs 0.2 hours for every researcher-hour it saves.
So a fully automated step that used to cost $95 of researcher time now costs 0.2 × $30 = $6 of AI time. The appendix writes this as (μ/w)·τ = 0.063. If every step in a field is automated and quality is held equal, output for a fixed budget rises by 1/0.063, or about 15.9. The main text says so directly: “a task that is fully automated by AI still costs about one-sixteenth of what it would cost a researcher to perform, so accelerations approach about 16x and stop there.”
The report is open about this. It says “halving the cost of AI, for example, would double the rate of progress,” and calls Scenarios 2 and 3 “illustrative scenarios, not forecasts.” What drops off when the number travels is that in the fields that reach 15x, the result is almost all assumption. The question “how much faster could computer science go?” has been answered by “how many dollars of agent time replace one dollar of researcher time?”, and the inputs to that are a list price, a salary survey, and a 0.2 the appendix says no data source measures.
Physical work does real work in the model, and that part I find more informative. In biology, chemistry, materials science, agricultural and biomedical sciences, physical steps are 38% to 41% of researcher time, against 4% in the median field, and full automation of everything else gets those fields to only 3x to 4x. That number comes from SciNet’s step classification, not from a price. It is the most robust claim in the scenarios: if a lab still has to pipette, the pipetting sets the pace.
The model also leaves out things the authors list themselves: field-level limits such as shared facilities, the cost of equipment and reagents, and the quality of ideas, held at q = 1. Gao and Wang’s 2024 study in Nature Human Behaviour, which measured AI use and benefits across the sciences, also found a “substantial gap between AI education and its application in research” and uneven benefits across disciplines. Diffusion takes time, and the report says that too.
What I’d take from it
The session data are worth having. At a scale nobody outside a lab can match, they show scientists bringing Claude data analysis and writing and rarely ideation, and computer science using Claude Code at seven to eight times its share of researchers. A specialized product also moves the mix: on Claude Science, ideation is 8% of sessions, roughly in line with the time it takes, against 4.2% on Claude.
The explanatory claim is weaker than its phrasing. “Use tracks capability” here means one model’s task labels, divided by another model’s time estimates, line up with a third model’s sense of what AI can do. That agreement is plausible and probably partly true. Without an outcome measure, it can’t be told apart from shared expectations.
And the 15-fold headline is a calibration, not a finding. It’s $95 over $6. Someone who thinks agent hours will be cheaper, or that agents will take longer per hour saved, can move it anywhere they like, and the model will follow. This fits a pattern I noticed in OpenAI’s internal agent-spend figures and in METR’s assessment of Anthropic’s own R&D speedup, published the same day as this report: the input numbers are getting more precise much faster than the output numbers.
Two things would change my reading. One is any outcome check on the capability score, even a small sample of tasks scored by researchers in the field. The other is a measured value for τ, the agent time per researcher-hour saved, from real sessions on a product that logs both. The first would tell us whether “use tracks capability” is about capability. The second would turn the ceiling from a price ratio into an estimate.
References
- Hermle, J., Jones, B., Althoff, L., Heller, R., Lyttelton, T., Stein, C., Wilmers, N., Massenkoff, M. and McCrory, P. (2026). AI for scientific discovery: How researchers use Claude. Anthropic, 5 November 2026.
- Hermle, J. et al. (2026). Appendix to “AI for Scientific Discovery”. Anthropic, 5 November 2026.
- Alamos, C. and Althoff, L. (2026). SciNet: The Anatomy of Science. Dataset, 2026.
- Liang, W., Zhang, Y., Wu, Z., Lepp, H. et al. (2025). Quantifying large language model usage in scientific papers. Nature Human Behaviour, 9(12), pp. 2599-2609. Published online 4 August 2025.
- Gao, J. and Wang, D. (2024). Quantifying the use and potential benefits of artificial intelligence in scientific research. Nature Human Behaviour, 8(12), pp. 2281-2292. Published online 11 October 2024.
- Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), pp. 1306-1308, 21 June 2024.
- Jones, B. F. (2009). The Burden of Knowledge and the “Death of the Renaissance Man”: Is Innovation Getting Harder?. Review of Economic Studies, 76(1), pp. 283-317, January 2009.
- Wuchty, S., Jones, B. F. and Uzzi, B. (2007). The Increasing Dominance of Teams in Production of Knowledge. Science, 316(5827), pp. 1036-1039, 18 May 2007.