Article Overview
Most AI tools were built for business. Claude Science was built for the lab.
Launched June 30, 2026, it is Anthropic's most ambitious expansion into scientific research — a fully integrated AI workbench that connects more than 60 scientific databases, manages compute resources from a single laptop to hundreds of GPUs, produces auditable reproducible figures, and orchestrates multiple AI agents working together on the same experiment.
The early results are hard to dismiss. A neuroscientist at the Allen Institute who previously needed two years to write a comprehensive literature review now completes them in a fraction of that time — with citation-checking agents verifying every source. An epidemiologist at UCSF confirmed Claude Science produces analyses as robust as manual methods while running them in roughly one-tenth the time. Manifold Bio used it to nominate drug targets end-to-end, integrating publicly available biology data with its own proprietary research history in a way no general-purpose coding assistant had previously managed.
This article covers what Claude Science is, how it actually works across its four major capabilities, what real scientists have done with it, and how to access it — including a grant program offering up to $30,000 in credits for research projects applying before July 15.
Introduction
Scientific research has a software problem that nobody talks about loudly enough.
A working biologist on any given day might pull data from UniProt for protein sequences, check PDB for structural information, query GEO for gene expression datasets, run a pipeline in Jupyter, submit a job to an HPC cluster, wait for it to finish, pull the results back, reformat them in R, and then start writing in a completely separate application. Each tool speaks a different language. Each database has its own query format. Each transition is friction — time spent not on the science, but on the logistics of doing science.
This is not a new problem. It is just one that has never had a satisfying answer until recently.
On June 30, 2026, Anthropic launched Claude Science — an AI workbench that brings these fragmented tools into a single environment where researchers can conduct all stages of their work. It is available in beta today for Claude Pro, Max, Team, and Enterprise users, and the early results from researchers who used it before public launch tell a specific, concrete story about what changes when the logistics disappear.
Quick Summary
| Detail | Information |
|---|---|
| Product | Claude Science |
| Launched | June 30, 2026 |
| Status | Beta |
| Available on | macOS and Linux |
| Plans | Claude Pro, Max, Team, Enterprise |
| Curated domain skills | 60+ (genomics, single-cell, proteomics, structural biology, cheminformatics) |
| Key partner | NVIDIA BioNeMo (Evo 2, Boltz-2, OpenFold3) |
| Compute partner | Modal (on-demand GPU scaling) |
| Grant program credits | Up to $30,000 per project (applications close July 15, 2026) |
| Access URL | claude.com/science |
The Problem That Made Claude Science Necessary
Understanding why Claude Science matters requires understanding what a typical research workflow actually looks like — not the idealized version, but the real one.
A researcher studying how germline genetic variants influence cancer susceptibility might need to pull variant data from ClinVar, cross-reference expression data in GEO, check pathway information in Reactome, validate protein function in UniProt, run association analyses in R, submit a compute job to a cluster for the heavy statistical work, wait for results, and then begin integrating all of it into a coherent picture they can eventually write up. Each of these steps is technically manageable in isolation. Together, in the sequence required by actual research, they constitute an enormous ongoing administrative burden that has nothing to do with scientific thinking.
Anthropic describes this as what makes scientific research "often tedious" — not the intellectual challenge of the science itself, but the constant translation between incompatible systems, tools, and data formats. Claude Science is built around the premise that this friction should not exist, and that eliminating it changes not just how fast science gets done, but what science becomes possible to attempt.
How Claude Science Actually Works
A Single Environment for the Entire Research Stack
Like Jupyter Notebook in its portability, Claude Science runs wherever researchers already work — locally on macOS or Linux, connected to a remote machine over SSH, or logged into an HPC node at a university computing cluster. The goal is not to replace the computing infrastructure a lab already relies on but to make that infrastructure accessible through a unified interface rather than a collection of separate tools.
At the center of that interface is a coordinating agent with access to more than 60 curated skills and connectors pre-configured for the most common research domains in biology: genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. This agent can spin up specialist sub-agents for specific tasks and can also interact with custom agents that researchers build themselves. The architecture is designed to grow with a lab's specific workflow rather than requiring researchers to adapt their methods to the tool.
Reproducible Artifacts That Carry Their Own History
Science only works if other people can verify what you did. This is the standard that separates peer-reviewed research from everything else — reproducibility. Claude Science takes this seriously in a way that most AI tools do not.
When Claude Science generates a figure, it does not just produce the image. It produces the image alongside the exact code that created it, the computational environment in which that code ran, a plain-language description of how the figure was built, and the complete message history of the session that produced it. Months after a figure was first generated, anyone with access to that artifact can trace every step from raw data to final output.
The same principle applies to manuscripts. Claude Science generates written scientific content alongside the evidence and processes that support it, with citations that can be traced back to their sources.
Beyond generation, the tool accepts plain-language editing instructions. Tell it to remove gridlines from a figure, change an axis to logarithmic scale, or adjust a color scheme — and it edits its own underlying code to produce the updated version. This keeps the code-figure relationship intact through every revision rather than allowing the two to diverge as they typically do when a researcher manually adjusts exported images in a separate application.
A reviewer agent runs throughout the process, checking citations against their sources, flagging any numbers that cannot be traced to their underlying data, and identifying figures whose visual output does not match what the code should theoretically produce. When it finds a problem, it flags it and self-corrects rather than waiting for a human to catch the discrepancy.
Compute Management That Scales From Laptop to Cluster
Large scientific analyses have their own logistics problem layered on top of the tool fragmentation problem. Running a genomics pipeline over a massive dataset, or folding a protein with a structure prediction model, requires a researcher to write and submit a job to a computing cluster, monitor whether it succeeded, retrieve the results if it did, and troubleshoot it if it did not — all of which pulls attention away from the actual science.
Claude Science handles this end to end. It drafts an execution plan, asks for permission before reaching new computing resources, and allows the researcher to review or revoke any decision before it is committed. It connects to a lab's existing HPC cluster over SSH or to Modal for on-demand cloud compute — scaling from a single GPU to hundreds as the analysis requires.
Three things make this approach genuinely useful rather than just convenient. First, large datasets only need to be loaded once because the agent session holds context in memory across the entire analysis — no repeated data transfers as the pipeline moves through stages. Second, sensitive datasets never leave the systems they already live on, because Claude Science runs on the lab's own infrastructure rather than uploading data to an external cloud. Only the context needed for each step of the analysis is sent to Claude. Third, researchers can fork a session at any point to compare two analytical approaches without losing the thread of the original — the equivalent of a save-state that works across an entire multi-hour research workflow.
Domain-Ready on Day One
Building a research pipeline from scratch typically means identifying the right databases for a given question, figuring out how to query each one, writing parsers for the various data formats they return, and integrating the results into whatever analysis environment the lab uses. This takes time even for experienced researchers who know the domain well.
Claude Science arrives pre-configured for the most common biological research workflows. When a researcher asks a question in plain language, specialist agents query and synthesize across databases including UniProt (protein sequences and function), PDB (protein structures), Ensembl (genomics), Reactome (biological pathways), ClinVar (genetic variants), ChEMBL (bioactive molecules), and GEO (gene expression data), along with journals and preprint servers.
The integration with NVIDIA's BioNeMo Agent Toolkit gives Claude Science native access to advanced life sciences models — Evo 2 for genomic sequence modeling, Boltz-2 for biomolecular structure prediction, and OpenFold3 for protein structure prediction. Researchers who already rely on these models as part of their validated workflow can access them through Claude Science without switching contexts.
Any pipeline a researcher builds in Claude Science can be saved as a reusable skill. Future sessions automatically inherit those skills, which means the customization work of the first session compounds over time rather than being rebuilt from scratch each time.
What Real Scientists Have Done With It
The three case studies Anthropic shares from the beta period are worth looking at carefully, because they are specific enough to evaluate rather than vague enough to dismiss.
Manifold Bio — Target Nomination for Tissue-Targeting Medicines
Manifold Bio designs medicines that home to a specific organ or cell type rather than distributing throughout the body — a precision approach that requires selecting the right molecular targets from among vast numbers of candidates. Their experiments test how millions of candidate binders corresponding to hundreds of targets distribute through a living body at once.
Claude Science was used to nominate the targets for Manifold's latest round of experiments. For each tissue and target combination, it assessed surface expression patterns, cellular trafficking behavior, and safety considerations — then ranked the candidates against criteria derived from Manifold's own internal research history on past programs.
Manifold drew a clear distinction between Claude Science and a general-purpose coding assistant: the difference was end-to-end completion. A coding assistant helps write the code for individual analytical steps. Claude Science gathered the right data, applied scientific judgment about what that data meant, integrated the context of Manifold's proprietary past programs, and delivered a ranked list of candidates ready for experimental follow-up — without requiring human intervention to bridge the gaps between steps.
Jérôme Lecoq — From Two Years to Weeks at the Allen Institute
Jérôme Lecoq is a neuroscientist at the Allen Institute who needed a way to write comprehensive scientific literature reviews — the kind that synthesize thousands of papers into a coherent, quantitative account of a research field. Before Claude Science, producing such a review could take his team up to two years.
Using Claude Science, Lecoq built a multi-agent "computational review template" comprising approximately 20 custom skills. Sub-agents read through thousands of papers, extracting the central claim and key quantitative finding from each one and storing them in a structured evidence database. A pipeline then constructed a narrative arc for the review, delegating each section to a specialized sub-agent. Within each section, dedicated agents generated quantitative cross-study figures drawn directly from the evidence database rather than described from memory.
A critical component of the workflow — specifically enabled by Claude Science's actor-critic architecture — was using paired agents for quality control: one agent created content while a separate reviewer evaluated it for accuracy and citation fidelity simultaneously.
The results: Lecoq's team now has approximately 10 completed reviews, many exceeding 100 pages, with citations verified by reviewer agents throughout. The same work that once took two years now takes a fraction of that time, and the team is continuing to refine the AI-based reviewer agents in collaboration with domain experts.
Stephen Francis — 10x Acceleration at UCSF on Glioma Research
Stephen Francis is an associate professor and epidemiologist at the UCSF Brain Tumor Center. His lab investigates glioma — a type of primary brain tumor that begins in glial cells — specifically focusing on how thousands of small-effect germline genetic variants combine across individuals to shape susceptibility to the disease.
This kind of research requires comprehensive germline analysis across multiple statistical approaches simultaneously. Before Claude Science, that work took the time it took. With Claude Science, Francis's lab now completes the same comprehensive workups in roughly one-tenth the time it previously required.
Critically, Francis's group did not take the tool's output on trust. They independently validated Claude Science's results against their manual methods and confirmed that the platform produces analyses that are not just faster but equally robust. Speed without accuracy would not be a useful research tool. Speed with accuracy changes what a research program can accomplish in a given year.
The AI for Science Grant Program
Anthropic is funding up to 50 Claude Science AI for Science projects with credits of up to $30,000 per project, alongside up to $2,000 in Modal compute credits for select projects.
The program is looking for research that spans scientific domains and pushes at the boundaries of what is currently possible, with an early focus on biology and biomedical research. Applications are open through July 15, 2026, with notifications sent by July 31. Funded projects run from September 1 through December 1, 2026.
For academic institutions and nonprofit research organizations, Anthropic has also introduced a discounted Team plan specifically designed for active scientific labs — a recognition that research institutions operate on different budget structures than commercial enterprises.
Getting Started
Claude Science is available in beta today at claude.com/science for users on Pro, Max, Team, and Enterprise plans. It runs on macOS and Linux — locally, over SSH to a remote machine, or through an HPC login node.
Team and Enterprise users require their administrator to enable Claude Science before it appears in their workspace. Individual Pro and Max users can access it directly.
The beta is being released early specifically to gather feedback from researchers using it on real problems. Anthropic describes the goal as building the platform alongside its scientific users rather than handing down a finished tool — a philosophy that the Allen Institute's iterative refinement of critic agents and Francis's independent validation work both reflect in practice.
For researchers who want to stay current on updates and connect with others in the community, Anthropic has launched an AI for Science Discourse forum for product announcements, feedback, and shared workflows.
Why This Matters Beyond the Features
The individual capabilities of Claude Science — the reproducible artifacts, the compute management, the 60+ curated connectors — are each useful on their own. But the more consequential claim in this announcement is the compound effect of putting them together.
A researcher who does not have to context-switch between PubMed and Jupyter and R and a cluster terminal is a researcher who can spend more of their thinking on the science itself. A review that used to take two years and now takes weeks is a review that gets written — and then another, and then another. An analysis that takes one-tenth the time does not just save nine-tenths of the timeline. It makes it economically possible to run ten analyses where before you could only run one, exploring ideas that would never have been attempted under the old constraints.
Anthropic's broader argument, stated plainly in the announcement, is that AI has the potential to dramatically accelerate the pace of scientific discovery and the development of healthcare interventions. Claude Science is their most concrete attempt yet to deliver on that potential — not through a general-purpose chatbot that can discuss biology, but through a purpose-built environment that thinks the way biological research actually works.
Final Takeaway
Claude Science is not an AI assistant that happens to know some biology. It is a research environment built from the ground up for how scientists actually work — fragmented tools unified, compute managed automatically, every output auditable and reproducible, and a reviewer agent watching for errors the researcher might miss.
The beta cases make the value concrete: two years to weeks on literature reviews, one-tenth the time on germline analysis, end-to-end drug target nomination that no general coding assistant could previously complete. These are not hypothetical improvements. They are documented results from real researchers at real institutions who independently validated what the tool produced.
For scientists whose work has been slowed as much by the logistics of research as by the difficulty of the science itself, Claude Science is the tool that removes the logistics. What researchers do with the time that frees up is entirely up to them.
