Anthropic has launched Claude Science, a beta AI workbench for researchers that brings scientific databases, code execution, compute management, artifacts, and domain-specific agents into one environment. The app is available now for Claude Pro, Max, Team, and Enterprise users on macOS and Linux, with Team and Enterprise access controlled by administrators.
The launch is not a new Claude model. It is a product bet that scientists will want AI inside the research workflow itself: reading papers, querying databases, writing and running analysis code, managing high-performance compute jobs, rendering scientific figures, and leaving enough traceability behind that a result can be checked later. That makes Claude Science a different kind of AI-for-science push than the model launches from OpenAI and Google DeepMind. Anthropic is trying to turn Claude into a coordinating layer around the tools labs already use.
In its announcement, Anthropic said Claude Science integrates common research tools and packages, produces auditable artifacts, and can access local machines, remote systems over SSH, or high-performance computing login nodes. Researchers interact with a generalist coordinating agent that can call more than 60 curated skills and connectors for areas including genomics, single-cell analysis, proteomics, structural biology, and cheminformatics.
What Claude Science Actually Does
The most useful detail is that Claude Science is built around reproducible work products, not just chat responses. Anthropic says generated figures include the code and environment that created them, a plain-language explanation, and the full message history behind the result. A researcher can ask for changes in ordinary language, such as changing an axis scale or adjusting a figure, and the system edits the underlying code rather than only altering an image.
The app also handles compute orchestration. For large jobs, such as protein folding or genomics pipelines, Claude Science can draft a plan, ask before reaching new resources, and prepare a job for the lab’s own infrastructure or for Modal compute. Anthropic says sensitive or large datasets can stay on the systems where they already live, with only the context needed for a step sent to Claude.
That design addresses one of the main frictions in computational biology: the work is scattered across literature databases, notebooks, R scripts, cluster terminals, genome browsers, structural viewers, and bespoke lab pipelines. Claude Science’s pitch is that a researcher can keep the scientific question in view while agents handle more of the plumbing between those tools.
NVIDIA Gives The Launch A More Technical Edge
The NVIDIA integration is the part that makes the launch more than another AI workspace announcement. NVIDIA said Claude Science can use its BioNeMo Agent Toolkit, which packages accelerated life-sciences capabilities as callable skills. Those skills can route work to models, libraries, and NVIDIA NIM microservices for tasks such as genomic analysis, protein-structure prediction, molecule design, and cheminformatics.
NVIDIA lists Evo 2, Boltz-2, and OpenFold3 among the scientific models and workflows that BioNeMo can expose to Claude Science. It also points to Parabricks for genomic analysis, RAPIDS-singlecell for single-cell workflows, and nvMolKit for cheminformatics operations such as similarity search and conformer generation. NVIDIA says 18 of the top 20 pharmaceutical companies already use BioNeMo, which gives Anthropic an enterprise life-sciences channel it would be hard to build from scratch.
The broader implication is that scientific AI agents will be judged by their tool access as much as their reasoning. A lab does not only need a model that can summarize papers. It needs an agent that can call a trusted docking workflow, query UniProt or ChEMBL, run a single-cell preprocessing job, produce a figure with auditable code, and hand a human researcher enough evidence to decide whether the next experiment is worth doing.
How It Compares With OpenAI And Google
Claude Science arrives after OpenAI introduced GPT-Rosalind, a life-sciences reasoning model for biology, drug discovery, protein engineering, and genomics. OpenAI later described a trusted-access structure for eligible organizations, emphasizing model capability and controlled deployment for research use.
Google is approaching the same market from the opposite direction: it owns major scientific AI assets. At I/O 2026, Google described Gemini for Science and Science Skills that integrate more than 30 life-sciences databases and tools, including UniProt, the AlphaFold Database, the AlphaGenome API, and InterPro. That gives Google a strong position in proprietary scientific models and reference resources.
Anthropic’s route is broader subscription access plus workflow integration. Claude Science is available to regular paid Claude tiers, not only a narrow set of enterprise research partners. That could make it attractive to academic labs and smaller research groups, especially if Anthropic’s discounted Team plan for active scientific labs is easy to adopt. It also means institutions will need to decide whether a generalist AI workbench is ready for regulated or sensitive research workflows.
Where Labs Should Be Careful
Claude Science’s most important promise is also its biggest governance question: it can sit close to code, data, compute, and scientific judgment. A reviewer agent that checks citations, calculations, and figure-code consistency is useful, but it is not the same as validated scientific review. Anthropic’s own case studies describe researchers independently checking results, which should remain the default expectation.
Labs considering the beta should start with low-risk workflows and clear boundaries. Good early use cases include literature triage, manuscript figures from non-sensitive data, reproducible analysis notebooks, pipeline documentation, and internal comparison of analysis approaches. Higher-risk use cases, such as clinical data, regulated drug-development decisions, proprietary assays, or dual-use biological work, need institutional review, data controls, and audit requirements before they move into an AI agent environment.
- Check whether the lab’s data can remain local or on approved HPC infrastructure.
- Confirm who can enable Claude Science on Team or Enterprise plans.
- Decide which connectors, SSH targets, databases, and compute accounts the agent may access.
- Require saved code, environments, prompts, and source references for any output used in research decisions.
- Keep human review mandatory for citations, statistical claims, biological interpretation, and experimental design.
Anthropic is also offering support for up to 50 Claude Science AI for Science projects, with up to $30,000 in Claude credits and up to $2,000 in Modal compute for select projects. Applications are open through July 15, 2026, notifications are planned by July 31, and selected projects will run from September 1 to December 1.
Why The Launch Matters
The AI-for-science market is moving from impressive demos toward operational systems. The winning products will not be judged only by benchmark performance or paper summaries. They will need to connect to trusted tools, preserve provenance, fit into institutional compute environments, and make it easier for scientists to inspect and reproduce what an AI system did.
Claude Science is still a beta, and many labs will need hands-on evidence before relying on it for serious research. But the shape of the product is telling. Anthropic is not asking scientists to leave their workflows for a chatbot. It is trying to put agents inside the machinery of research itself.
