Anthropic released Claude Science on June 30, a multi-agent AI workbench built for researchers working in genomics, proteomics, structural biology, and cheminformatics. The product, now in beta for Claude Pro, Max, Team, and Enterprise subscribers, bundles over 60 pre-configured scientific skills and connectors into a single environment where specialized agents coordinate literature analysis, data pipeline execution, and figure generation with full provenance tracking.

How the Multi-Agent Architecture Works

Claude Science uses a coordinating generalist agent that receives plain-language requests and delegates to specialized sub-agents. According to Anthropic’s announcement, researchers interact with the coordinator while domain-specific agents handle established workflows for their field. A separate reviewer agent runs in parallel, inspecting outputs step by step: flagging incorrect citations, untraceable numbers, and figures that don’t match their underlying code, then self-correcting as the pipeline executes.

Every generated figure includes the exact code, environment, and full message history that produced it. Researchers can edit outputs in natural language (“change this axis to log scale”) and fork sessions to compare two analytical approaches without losing the original thread.

NVIDIA BioNeMo Integration

The workbench connects natively to NVIDIA’s BioNeMo Agent Toolkit, giving Claude Science agents access to GPU-accelerated life sciences models including Evo 2 (genomics), Boltz-2 (biomolecular interaction prediction), and OpenFold3 (protein structure prediction). Domain-ready connectors span UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO, and additional databases, as reported by MarkTechPost.

Compute scales on demand. Claude Science submits jobs to a lab’s existing HPC cluster over SSH or to Modal for cloud GPU access, scaling from a single GPU to hundreds. Sensitive datasets stay on local infrastructure because only the context needed for each analysis step is sent to Claude.

Early Results From Beta Users

Three beta users demonstrate the product’s scope:

Manifold Bio, a tissue-targeting drug design company, used Claude Science to nominate targets for its latest experiments. For each tissue and target, the workbench assessed surface expression, trafficking, and safety, then ranked candidates against Manifold’s proprietary criteria. The company told Anthropic that end-to-end execution distinguished Claude Science from general coding assistants.

Jérôme Lecoq at the Allen Institute built a multi-agent “computational review template” with roughly 20 custom skills. Sub-agents read thousands of papers, extracted central claims and key quantitative findings, stored them in an evidence database, then constructed narrative arcs section by section. Lecoq told Anthropic his team previously spent up to two years on a single comprehensive review. He now has about 10 reviews, many over 100 pages, with citations verified by reviewer agents.

Stephen Francis, an epidemiologist at the UCSF Brain Tumor Center, used Claude Science to accelerate germline variant analysis for glioma susceptibility research. Francis told Anthropic the app completed comprehensive workups across multiple approaches in roughly one-tenth the time his lab previously required. His group independently validated the results.

Pricing and Research Grants

Claude Science is available on macOS and Linux. Team and Enterprise users need admin approval to enable access. Anthropic is offering a discounted Team plan for academic labs and nonprofit research organizations.

The company will fund up to 50 Claude Science research projects with up to $30,000 in credits each, with cloud compute partner Modal providing up to $2,000 in additional compute per project. Applications close July 15, 2026, with projects running from September through December.

The Vertical Agent Bet

Claude Science is Anthropic’s clearest move toward vertical-specific agent deployments. Rather than shipping a general-purpose tool and letting researchers figure out scientific workflows, Anthropic pre-loaded domain knowledge, connected to established databases, and partnered with NVIDIA for GPU-accelerated biology models. The architecture, where a coordinator delegates to specialized sub-agents while a reviewer checks their work, mirrors production agent patterns emerging across enterprise software. The difference: Claude Science applies them to a domain where reproducibility and provenance are non-negotiable, and where compressing a two-year literature review into days has obvious commercial value.