Anthropic launched Claude Science, an AI workbench for scientific research, on June 30, 2026. (Photo by JOEL SAGET / AFP via Getty Images)
AFP via Getty Images
Today I handed an AI my entire research library, 6,576 papers, and asked it a question I had been chewing on for years. Do the formal vocabularies that are supposed to encode my field actually capture the way people in that field think?
I study how diseases are transmitted from animals to humans. Spillover science, like most fields, has official ontologies and curated catalogs of the concepts and relationships that are used to organize the work. My suspicion was that they were thin. So instead of arguing about it, I tested it.
The system read the 490 articles in my collection about the spread of zoonoses and the emergence of pandemics, pulled concepts and causal claims from the full text, and assembled a vocabulary from scratch. Then it contrasted this with the formal ontologies. The discrepancy was significant. Of 915 relationships used repeatedly in the literature, 864 had no counterpart in the standard reference. Twelve hundred conceptual categories appeared in only 490 papers and nowhere in the formal schemes, grouped by environmental factors and ecological processes. It turned out that the working language in just a fraction of the literature in my field was about four times larger than the official language.
The project cost $26 plus change.
The latent ontology of spillover science is much more extensive than the formal one. Learned de novo from the full text of 487 papers and compared with four formal ontologies, it contains 1,240 conceptual classes and 864 relational predicates with no formal counterpart, as well as a concept space that is approximately four times higher dimensional. (Author’s analysis, created with Claude Science)
JM Drake
This experiment was my test drive with Claude Science, the tool Anthropic released today. The company’s goal is to do for laboratory research what Claude Code did for software, and the ambition is no small one. Six months ago, Zubair Jandali, who leads healthcare and life sciences at Anthropic, told an audience that Claude could help with the digital work of life sciences. The pitch today was that it can do this work.
Table of Contents
The same model, a new harness
Start with what Claude Science is not. This is not a new model. Anthropic is unusually clear about this: the product runs on the same Claude that everyone already uses, including Opus 4.8, with no special access and no gating. The intelligence is the same intelligence that everyone can already rent.
The dishes around it are new. In artificial intelligence, a harness is the framework that turns a general-purpose model into a working tool, the connections to data, the ability to execute code, the memory of what it has done and the verification of its output. A model alone can reason about a protein. A model in a good system can retrieve the structure from a database, fold a variant into a cluster, render the result, and log each step. Claude Science is a tableware designed for science, and a substantial one at that.
It connects more than sixty scientific databases, has out-of-the-box genomics, proteomics, structural biology, and chemistry capabilities, renders protein structures and chemical diagrams inline, and manages computational tasks via a laptop, cluster, or rented GPUs. Each character it generates contains its complete history, code, computing environment, and conversation that generated it, bundled together so that the result can be regenerated later.
None of this makes the model smarter. It makes the model useful, which is the more valuable thing at the moment. A computational biologist with Claude Code and a GitHub account could assemble much of this himself with a few weeks of wiring. Claude Science is betting that doing this wiring once will beat a thousand labs doing it from scratch each time. The value lies in the curation, and curation is what turns raw skill into science.
Built for the bench, open for use
The launch demo was a drug discovery campaign. Starting from a single sentence, Claude planned and executed a search for a molecule to stabilize the broken enzyme behind phenylketonuria, screened 2,200 compounds on 80 GPUs, narrowed them down to four candidates, and created a go/no-go memo. The same triage was then performed for 100 rare diseases simultaneously. Why stop at 100, the presenter asked, when the same machine could just as easily run at 10,000?
It was an impressive achievement, and it is completely molecular. Every database, every ready-made capability, every partner model points to the same kind of science: genes, proteins, small molecules, structures. OpenAI and Google have aimed their own scientific tools at the same goal: pharmaceutical research, where the money is.
The rest of science, an enormous field, is wide open. The earth and atmospheric sciences, the environmental sciences, ecology, the social and behavioral sciences, and much of the epidemiology: none of these are configured yet, and the data on which these areas are based, biodiversity records, climate reanalyses, census files, remote sensing, are not among the sixty databases. Science that occurs outside the wet lab, in the field, in the watershed, and in the population is the next frontier for a tool like this.
You can easily imagine it. A harness like this, aimed at my field, could retrieve species occurrence records from GBIF, overlay them over climate reanalysis, fit a distribution model and mark the counties where a tick-borne pathogen is most likely to spread in the next season, and then design the monitoring mission and associated figure. It could read every outbreak report from a region and reconstruct the chain of transmission. It could do for a health department what the demonstration did for a drug program, condensing weeks of meetings into an afternoon. The intelligence for this already exists. What’s missing are the wiring harness, connectors, and capabilities, and my $26 experiment is a little proof that it can be done.
The engine is general. Anthropic aimed its first system at pharmaceutical research, where the urgency and budgets are large, but nothing about the technology limits them to that. Field science is an invitation, not an oversight.
Where scientists come into play
The technology is exciting and will accelerate much of science. It also highlights where scientists are most important. Once production is cheap, all that remains is judgment work. Auditing, validation and correction are those Actually rate-limiting steps of virtually all work that crosses a screen, and that is simply a description of what good science has always been. Claude Science even has a review agent who reports erroneous citations and mismatched numbers. It is the same model that checks its own work for now, not an independent source of truth, but the direction is right.
The question that I find most exciting is novelty. A model trained on the existing literature is exceptionally good at reproducing that literature, and the risk is regression toward a mediocre mean, a field that speaks to itself. But the same tool that maps what a field already believes can reveal its gaps, the relationships that no one has tested, the concepts that appear everywhere and are nowhere defined. For twenty-six dollars I found the edges of what my field had already written down. Finding what it hasn’t yet imagined is the harder and more exciting problem, and for the first time it looks like we might actually be able to tackle it.
https://www.forbes.com/sites/johndrake/2026/06/30/anthropics-new-ai-workbench-mapped-my-field-for-26-now-imagine-it-aimed-at-the-rest-of-science/
