Claude Science brings research, code and compute together

What this is about

Claude Science is not a general model announcement, but a concrete tool: Anthropic describes it as an AI workbench for scientific work, available since June 30, 2026. According to Anthropic, the beta is available for Claude Pro, Max, Team and Enterprise users on macOS and Linux.

The point is not that researchers get another chatbot. Claude Science tries to bring the many tools of a research day into one shared workspace: literature sources, Jupyter, R, terminals, specialist databases, protein and molecule views, manuscript work and compute jobs.

What Claude Science actually does

Claude Science is an app with a coordinating agent, more than 60 curated skills and connectors for areas such as genomics, single-cell analysis, proteomics, structural biology and cheminformatics. Anthropic names PubMed, Jupyter, R, HPC logins, Modal compute and scientific databases as typical parts of the environment it is meant to work with.

The artifact idea is important: when Claude Science creates a figure or analysis component, the environment is meant to carry along the code, compute environment, explanation and message history. That matters more in research than a polished answer, because results need to be validated, changed and reproduced later.

The app can run on local hardware, over SSH or on HPC infrastructure, according to Anthropic. Large or sensitive datasets therefore do not have to be copied wholesale into an external web interface; Anthropic says only the context needed for a given step is sent to Claude.

Why it matters

Research rarely fails only because a model is missing. It fails at handoffs: PDF to dataset, database to notebook, notebook to cluster, cluster result to figure, figure to manuscript, manuscript to citation check. Claude Science targets exactly those breaks.

For real users, this is interesting when their work depends on repeatable analyses: bioinformatics, drug discovery, literature reviews, structured evaluations or internal review processes. Anthropic cites examples in single-cell RNA sequencing, CRISPR screen design, protein structure prediction and cheminformatics.

The value is therefore less a magic answer and more an ordered workspace. The reviewer agent is especially relevant because it is meant to check citations, numbers and figures and trigger corrections. That does not replace expert review, but it can make traceability problems visible earlier.

In plain language

Imagine a research kitchen. Before, the recipe, ingredient list, scale, oven, notebook and shopping list were in different rooms. Claude Science tries to put everything on one worktop and write down which flour, temperature and baking time were used for every loaf.

A human still has to taste the bread. But if it works or fails, it is clearer later why.

A practical example

A small bioinformatics team analyzes 10,000 single-cell profiles from a study on inflammatory responses. Until now, a PhD student searches literature in PubMed, writes Python code in a notebook, starts compute jobs on a cluster and manually moves results into a manuscript.

With Claude Science, the team could create an analysis session: the agent gathers relevant papers, drafts a first notebook, starts a cluster job for normalization, creates three candidate figures and documents the code behind each figure. The reviewer agent flags two unclear sources and one number that does not match the table. The PhD student decides which corrections are scientifically valid and exports a manuscript section with traceable artifacts.

The gain would not be that truth automatically appears from 10,000 cells. The gain would be less time lost between tools and more steps that remain checkable.

Scope and limits

• Claude Science is in beta. Teams making regulatory or clinical decisions still need their own validation, approvals and audit process.
• The tool can connect research, code and compute, but it cannot replace wet-lab experiments, domain judgment or peer review.
• Privacy remains a project decision. Local or HPC-near operation helps, but teams must check exactly which data and context are sent to Claude.

A sensible first test is not a risky production process. A better test is a completed internal analysis project: same data, known target finding, and a clear measurement of whether Claude Science saves time and whether the artifacts are actually reproducible.

SEO & GEO keywords

Claude Science, Anthropic, AI research workbench, scientific AI tools, Jupyter AI, PubMed AI, HPC workflows, reproducible research, life sciences AI, research automation