Anthropic bets on Claude Science as a lab-grade AI workbench to speed reproducible science

Anthropic launched Claude Science in beta on June 30, 2026, an AI workbench that bundles a researcher's databases, code tools, and compute into one environment. A coordinating agent orchestrates sub-agents and a separate reviewer agent checks citations and math, aiming to make scientific analysis faster and more reproducible.

What happened

Claude Science is not a new model. It runs the same Claude models already on sale, including Opus 4.8, with no special access. The bet is on workflow rather than raw model power. The workbench sits on a lab's own infrastructure to keep data in-house, working locally on macOS or Linux, or on a remote box over SSH or an HPC login node. Large jobs, such as folding a protein or running a genomics pipeline, can be submitted to the lab's own cluster or to a Modal account for on-demand compute, scaling from one GPU to hundreds.

Anthropic integrated the Nvidia BioNeMo Agent Toolkit to access life-sciences models such as Evo 2, Boltz-2, and OpenFold3. The system also draws on more than 60 scientific databases, including UniProt, PDB, and ChEMBL. The coordinating agent can spin up specialist sub-agents built by the user. A reviewer agent then checks outputs for untraceable figures or mismatched references, flagging and correcting errors as it goes.

Anthropic pointed to three beta users. Manifold Bio used the app to nominate targets for experiments. Jerome Lecoq at the Allen Institute built a multi-agent template of about 20 custom skills to write long-form reviews. Stephen Francis at the UCSF Brain Tumor Center said his glioma analysis ran in about a tenth of the usual time, and the group confirmed the results by hand.

Why AI builders should care

Claude Science targets a real pain point: scientists work across dozens of databases, each with its own schema, switching between PubMed, Jupyter, R, and cluster terminals. The workbench folds those steps into one environment and lets an AI agent handle the orchestration. For AI builders, the product demonstrates how to embed models into domain-specific workflows with reproducibility as a first-class concern.

Every figure in Claude Science arrives with the exact code and environment that produced it, plus a plain-language note on how it was made and the full message history. A researcher can return months later and trace any result. This auditable artifact approach is directly applicable to any AI product that needs verifiable outputs.

Anthropic is also making a commercial push. The company is racing to win paying customers ahead of a planned listing, and funding up to 50 research projects with up to $30,000 in credits each. Applications are open until July 15, 2026.

Practical implications

Claude Science runs on the lab's own machines, so large or sensitive datasets never have to leave. Only the context needed for each step is sent to Claude. Researchers can fork a session to compare two approaches without losing the original. The design answers a privacy worry that often blocks AI adoption in research.

The tool also uses an agentic approach: the coordinating agent drafts a plan and asks before reaching new resources. This pattern is useful for building any AI system that needs to operate across multiple external tools and data sources while respecting cost or access boundaries.

Caveats

The product is in beta, and real-world iteration may alter support for certain workflows, databases, or models. Early coverage notes that Claude Science is primarily built for molecular biology and may overlook fields like ecology or epidemiology. Beta details come from announcements and media coverage; some capabilities may evolve.

A tool that can turn a two-year review into a batch of 10 long-form papers could accelerate synthesis but also flood the literature. Anthropic's mitigation is the reviewer agent and human checks. Builders adopting similar patterns should plan for safeguards around output quality and provenance.

FAQs

What is Claude Science and what does it do in the lab?

Claude Science is an AI workbench that consolidates researchers' databases, code tools, and compute under a coordinating agent to run end-to-end analyses. A separate reviewer agent checks citations and math to ensure reproducibility.

How does Claude Science consolidate researchers' databases and tools?

It brings together databases, code editors, and compute resources in one environment. The coordinating agent can spin up sub-agents and the reviewer agent validates outputs, ensuring traceability and reproducibility.

What databases and life-science models does Claude Science access?

Through the Nvidia BioNeMo Agent Toolkit, it accesses models like Evo 2, Boltz-2, and OpenFold3. It also pulls data from more than 60 databases including UniProt, PDB, and ChEMBL.

Can Claude Science run on in-house infrastructure and HPC clusters?

Yes. The workbench runs locally on macOS or Linux, or on a remote HPC via SSH. Large jobs can be submitted to the lab's own cluster or to Modal for on-demand compute, keeping data in-house.