Claude Science is Anthropic's new AI workbench built specifically for researchers, bundling literature analysis, multi-step data workflows and citation-checking into a single tool that can run entirely on a lab's own hardware. Anthropic says the app pulls together dozens of scientific databases, tools and software packages so researchers can analyze papers, run multi-step analyses, generate charts and draft manuscripts without switching between separate systems.
What Claude Science can do for research teams
The tool ships with more than 60 preconfigured skills covering fields such as genomics, proteomics and cheminformatics, according to Anthropic. A built-in verification agent automatically checks citations and calculations as it works, addressing one of the biggest complaints researchers have about AI-generated analysis: unverified or fabricated references. Anthropic frames Claude Science as a way to move an AI assistant from casual literature summaries toward the kind of rigor scientific publication actually demands.
Researchers can also save their own analysis pipelines as reusable skills, letting a lab standardize a workflow once and reuse it across future projects rather than rebuilding it from scratch every time a similar question comes up. That reusability matters most in labs juggling many parallel projects, where a postdoc or grad student would otherwise spend hours re-wiring the same literature search or data-cleaning steps for each new dataset.
Local and HPC deployment keeps lab data in-house
Claude Science runs locally on macOS or Linux and can connect out to remote machines through SSH or high-performance computing (HPC) clusters, which means sensitive research data never has to leave a lab's own infrastructure. Anthropic says only the context the model actually needs gets sent externally, and the app can scale a job from a single GPU up to hundreds when a workload demands it.
That local-first design matters for fields like genomics and drug discovery, where research data is often proprietary, tied to patient consent restrictions, or subject to institutional data-governance rules that make sending everything to a cloud service a nonstarter. For labs evaluating agentic AI tools, the ability to keep data on internal infrastructure while still tapping a capable model is likely to be a bigger factor than raw benchmark scores, since a single compliance violation can shut down a research program far faster than a slower analysis pipeline ever would.
Built on Nvidia's BioNeMo toolkit
Under the hood, Claude Science taps Nvidia's new BioNeMo agent toolkit, which ships with specialized models including Evo 2, Boltz-2 and OpenFold3 — tools built for tasks like protein structure prediction and genomic sequence modeling. Pairing a general-purpose AI assistant with domain-specific scientific models is an increasingly common pattern as AI vendors try to make agentic tools useful for scientific research rather than just general knowledge work.
Claude Science is available now in beta for Anthropic's Pro, Max, Team and Enterprise subscribers. Anthropic is also backing up to 50 research projects with as much as $30,000 in credits each, with applications open through July 15, 2026 — a push that suggests the company sees specialized, verifiable research tooling as a priority use case for agentic AI, not just a side project. As more labs weigh whether to adopt AI-assisted analysis, tools that verify their own output rather than asking scientists to take results on faith are likely to see faster buy-in from skeptical research teams.
Claude Science: quick answers
What is Claude Science used for?
It's an AI workbench for scientific researchers that combines literature analysis, multi-step data workflows, chart generation and manuscript drafting, backed by a verification agent that checks citations and calculations.
Who can access Claude Science, and what does it cost?
It's available now in beta at no extra cost to Anthropic's Pro, Max, Team and Enterprise subscribers. Anthropic is also offering up to $30,000 in research credits to as many as 50 projects, with applications open through July 15, 2026.
