Scientific research has always been slow, not because scientists lack ambition, but because the tools they rely on were never built to work together. A typical researcher might spend their day bouncing between PubMed, Jupyter notebooks, R, a cluster terminal, and half a dozen specialized databases, each with its own quirks. Anthropic thinks it has a better answer.
The company has launched Claude Science, an AI workbench that brings those fragmented tools into a single environment. It is available now in beta for Claude Pro, Max, Team, and Enterprise subscribers on macOS and Linux. The launch follows Anthropic’s entry into life sciences research last fall and represents its most significant push into the scientific market so far.
The timing matters. AI’s potential to speed up drug discovery and biomedical research has attracted intense interest from companies including Google DeepMind, Microsoft, and a wave of well-funded startups. Anthropic is positioning Claude Science as a practical, day-to-day research tool rather than a narrow solution for a single task, which is a meaningful distinction in a market that is still figuring out what useful AI looks like for scientists.
What Claude Science actually does
At its core, Claude Science is a coordinating agent that can call on more than 60 pre-configured skills and connectors covering genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. Users interact with it in plain language, and specialist agents handle the actual querying and synthesis across sources like UniProt, PDB, Ensembl, ClinVar, and ChEMBL.
A few things set it apart from a general-purpose coding assistant:
- It renders scientific outputs natively, including 3D protein structures, genome browser tracks, and chemical structures, directly in the interface.
- Every figure comes with the exact code and environment that produced it, a plain-language explanation, and a full message history, so results can be validated and reproduced months later.
- A built-in reviewer agent checks citations and calculations as the work progresses, flagging and correcting errors automatically.
- Users can edit figures in plain language, such as asking to remove gridlines or switch an axis to log scale, and the agent rewrites its own code accordingly.
Claude Science also connects to NVIDIA’s BioNeMo Agent Toolkit, giving it access to life sciences models including Evo 2, Boltz-2, and OpenFold3.
Compute management without the headaches
One of the more practical features is how Claude Science handles large computing jobs. Researchers who need to fold a protein or run a genomics pipeline over a massive dataset normally have to stop what they are doing, set up a job, send it to a cluster, wait, check whether it worked, and pull the results back. Claude Science takes over that process.
It drafts a plan, asks for confirmation before using new resources, and submits jobs to whatever infrastructure the lab already uses, whether that is an HPC cluster over SSH or a Modal account for on-demand compute. It can scale from a single GPU to hundreds as needed.
Because agents work inside a running session that holds context in memory, large datasets only need to be loaded once. Sensitive data stays on the lab’s own systems, and only the context needed for each step is sent to Claude’s servers.
Early results from researchers in the field
Anthropic has been running a private beta for several months. A few examples stand out.
Manifold Bio, which designs tissue-targeting medicines, used Claude Science to select targets for its latest experiments. The app assessed surface expression, cell trafficking, and safety data for each candidate, then ranked them against criteria drawn from Manifold’s internal data. The company said the end-to-end capability, pulling the right data and applying judgment with the context of past programs built in, was what distinguished it from a general coding assistant.
Neuroscientist Jerome Lecoq at the Allen Institute built a multi-agent pipeline using roughly 20 custom skills to write long-form scientific reviews. Sub-agents read thousands of papers, extracted central claims and key quantitative findings, and stored them in an evidence database. Separate agents then built a narrative structure, wrote each section, and generated cross-study figures directly from that database. Actor-critic pairs, where one agent creates content and another checks it for accuracy and citation fidelity, were central to the workflow. Writing a single review used to take his team up to two years. He now has around 10 reviews, many exceeding 100 pages.
Stephen Francis, an associate professor and epidemiologist at the UCSF Brain Tumor Center, used Claude Science to support studies on the molecular epidemiology of glioma. His lab looks at how thousands of small-effect genetic variants combine to shape individual cancer risk. He said Claude Science reduced the time needed for comprehensive genetic analysis to roughly one-tenth of what it previously took, and his team independently validated the results.
Pricing, access, and grant funding
Claude Science is available in beta on macOS and Linux for Pro, Max, Team, and Enterprise plans. Team and Enterprise users need an admin to enable it. Anthropic has introduced a discounted Team plan for active scientific labs at academic institutions and nonprofit research organizations.
For researchers who want financial support, Anthropic is offering up to 50 AI for Science project grants, each worth up to $30,000 in credits. Modal is adding up to $2,000 in compute for selected projects. The program is focused early on biology and biomedical research, though other domains are welcome to apply. Applications close July 15, 2026, with awards announced by July 31. Projects run from September 1 to December 1, 2026.
Anthropic has also set up an AI for Science Discourse community for researchers to share feedback and learn from each other. The app is available at claude.com/science.