Scientific research has always been slow, fragmented, and expensive. Researchers bounce between dozens of databases, wrestle with incompatible file formats, and spend more time setting up computing jobs than actually doing science. Anthropic thinks AI can fix that, and it’s putting that bet into a new product.
Anthropic announced Claude Science on June 30, 2026, describing it as an AI workbench for scientists. The app is now available in beta for Claude Pro, Max, Team, and Enterprise users on macOS and Linux. It’s the company’s biggest push yet into scientific research, following the life sciences work it started last fall.
The timing matters. AI companies are racing to show that their models can do more than write emails and summarize documents. Science is one of the hardest tests, and also one of the most valuable. If AI can meaningfully speed up drug discovery, genomics research, or cancer epidemiology, that’s a case study no other industry can match.
What Claude Science actually does
At its core, Claude Science is a research environment that tries to replace the patchwork of tools scientists currently use. Instead of switching between PubMed, Jupyter notebooks, R, and a cluster terminal, researchers work inside a single app connected to all of them.
The system runs on a coordinating AI agent backed by more than 60 pre-configured skills and connectors, organized around the most common research areas:
- Genomics and single-cell analysis
- Proteomics
- Structural biology
- Cheminformatics
The main agent can spin up specialist sub-agents as needed, and a separate reviewer agent runs in the background checking citations, calculations, and whether figures actually match the code that generated them. That last part is a direct response to one of the biggest criticisms of AI-generated research: you can’t always trace where a number came from.
Reproducibility built into every output
Reproducibility is a serious problem in science generally, and AI tools have made it worse in some cases by producing outputs with no audit trail. Claude Science takes a different approach.
Every figure or manuscript the app generates comes with the exact code that produced it, the computing environment at the time, a plain-language explanation of how it was created, and the full message history. If a researcher needs to validate a result six months later, everything they need is already attached to the output.
Scientists can also edit figures by describing what they want in plain language. Asking to remove gridlines or switch an axis to log scale causes the agent to rewrite its own code, so the change is tracked rather than applied as a one-off tweak.
Compute management without the headaches
Running large analyses has always required scientists to step away from their actual work. Folding a protein or processing a big genomics dataset means writing a job submission script, waiting for a cluster queue, checking whether it ran, and pulling the results back. Claude Science handles that entire process.
The app works with infrastructure labs already have, including HPC clusters via SSH and Modal for on-demand GPU compute. It drafts a plan before doing anything, asks before accessing new resources, and lets researchers review or cancel any step. Analysis can scale from a single GPU to hundreds depending on what the job requires.
Because agents run inside a persistent session, large datasets only need to load once. The data stays on the lab’s own infrastructure, so sensitive or proprietary datasets don’t have to move to a third-party server. Only the context needed for each step goes to Claude.
Connected to the scientific databases scientists actually use
One of the more practical features is how Claude Science handles data sources. Biological research data is scattered across resources like UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, and GEO, each with its own query language and schema. Claude Science connects to all of them, so a plain-language question pulls from multiple sources without the researcher having to query each one manually.
The app also uses NVIDIA’s BioNeMo Agent Toolkit to connect to life sciences models including Evo 2, Boltz-2, and OpenFold3. Labs can save their own pipelines as reusable skills, connect their preferred tools with a custom connector, and have future sessions inherit those configurations automatically.
What researchers have done with it so far
Anthropic has been running a beta and shared three specific examples from that period.
Manifold Bio, which designs tissue-targeting medicines, used Claude Science to nominate targets for its latest experiments. For each tissue and target, the app assessed surface expression, trafficking, and safety, then ranked candidates using criteria from Manifold’s own internal data. According to Manifold, what separated it from a general coding assistant was the ability to run the full process end-to-end, gathering the right data and applying judgment with context from past programs already built in.
Jerome Lecoq, a neuroscientist at the Allen Institute, built a multi-agent pipeline with about 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. The pipeline then constructed a narrative arc for the review, with dedicated agents generating quantitative figures directly from that database. A critic agent evaluated each section for accuracy and citation fidelity. His team previously needed up to two years to write a review of this kind. He now has about 10 reviews, many over 100 pages, with citations checked by reviewer agents.
Stephen Francis, an associate professor and epidemiologist at the UCSF Brain Tumor Center, used Claude Science for research on the molecular epidemiology of glioma. His lab studies how thousands of small-effect genetic variants combine to shape cancer risk. Claude Science let his team run comprehensive analyses across multiple approaches in roughly one-tenth the time it previously took. His group independently validated the results.
Pricing, availability, and research grants
Claude Science is available now in beta on macOS and Linux for Pro, Max, Team, and Enterprise subscribers. Team and Enterprise users need an admin to enable it. Anthropic has a discounted Team plan for scientific labs at academic institutions and nonprofit research organizations.
For researchers who want to go further, Anthropic is offering up to 50 funded projects through its AI for Science program:
- Up to $30,000 in Claude credits per project
- Up to $2,000 in compute credits from Modal for selected projects
- Focus on biology and biomedical research, with other domains welcome
- Applications open through July 15, 2026
- Award notifications by July 31, with projects running September 1 to December 1, 2026
Researchers can get started at claude.com/science and join the AI for Science Discourse community for updates and feedback.
