Anthropic's Claude Science Maps Research Fields for $26

A $26 experiment reveals AI's research potential

A University of Georgia researcher fed 490 papers on zoonotic spillover into Anthropic's new Claude Science tool and discovered something striking: the working vocabulary scientists actually use in his field is four times richer than the formal ontologies meant to organize it. Of 915 relationships the literature referenced repeatedly, 864 had no counterpart in standard reference schemes. The analysis cost $26 and took hours, not months.

The experiment, conducted by professor John Drake, served as a real-world test of Claude Science, which Anthropic released on June 30, 2026. The tool represents the company's effort to do for laboratory research what its Claude Code product did for software development.

Why it matters

Claude Science demonstrates how AI can compress weeks of research synthesis into affordable, rapid analysis—but it also exposes a critical gap. While Anthropic and competitors like OpenAI and Google have aimed their scientific tools at pharmaceutical research, vast domains including ecology, epidemiology, environmental science, and field research remain unconfigured. The underlying technology is general-purpose; what's missing is the connective tissue to make it useful beyond the wet lab.

Not a new model, but a new harness

Claude Science is not a more advanced AI model. It runs the same Claude Opus 4.8 available to all users. What Anthropic built is what the AI industry calls a "harness"—the scaffolding that transforms general intelligence into a working scientific tool.

The harness connects Claude to more than 60 scientific databases and includes pre-built capabilities for genomics, proteomics, structural biology, and chemistry. It can render protein structures and chemical diagrams, manage computing jobs across different hardware configurations, and maintain full provenance for every result it generates. Each figure carries its complete history: the code, the computing environment, and the conversation that produced it.

According to details first reported by Forbes, the launch demonstration focused on drug discovery. The system planned and executed a search for molecules to stabilize a broken enzyme linked to phenylketonuria, screening 2,200 compounds across 80 GPUs and narrowing results to four candidates. It then replicated this process across 100 rare diseases simultaneously.

The field sciences remain wide open

Every database and pre-built skill in Claude Science currently points toward molecular research—genes, proteins, small molecules, and structures. This reflects where pharmaceutical money concentrates, but it leaves enormous scientific domains unaddressed.

Earth sciences, atmospheric research, ecology, behavioral sciences, and much of epidemiology operate on entirely different data: biodiversity records, climate reanalysis, census files, and remote sensing. None of these appear among the 60 integrated databases. A harness configured for field sciences could pull species occurrence records, overlay climate data, model disease vector distributions, and draft surveillance briefs—compressing weeks of work into hours. The intelligence exists; the connectors do not.

Where human judgment remains essential

As generation becomes cheap, validation becomes the rate-limiting step. Claude Science ships with a reviewer agent that flags questionable citations and numerical mismatches, though it currently uses the same model to check its own work rather than an independent verification source.

The deeper challenge is novelty. A model trained on existing literature excels at reflecting that literature back, which risks reinforcing what a field already believes rather than identifying what it hasn't yet imagined. The same tool that maps established knowledge can expose gaps—relationships untested, concepts undefined—but finding genuinely new questions remains the harder problem.

Drake's $26 analysis revealed the edges of what his field has documented. The more valuable frontier, he notes, is discovering what it hasn't yet conceived.

Details of the Claude Science launch and Drake's experiment were first reported by Forbes.