Anthropic Launches Claude Science, an AI Workbench for Scientific Research

Anthropic has launched Claude Science, an AI workbench that brings together tools and packages scientists rely on, produces outputs with a clear audit trail, and offers flexible access to computing power. This approach is part of a broader trend toward orchestration, where orchestration beats raw model size.

Fragmented Workflows, Unified Environment

Scientific research often means hopping between different databases, file formats, and tools like PubMed, Jupyter, R, and cluster terminals. Claude Science pulls these pieces into a single environment, letting scientists handle literature analysis, multi-step research, figure and manuscript polishing from one place. Every output comes with an auditable history for validation and reproducibility. It can be run locally on macOS or Linux, or accessed remotely via SSH or an HPC login node.

Agent-Driven Collaboration

Users work with a generalist coordinating agent that taps into over 60 curated skills and connectors, pre-configured for fields such as genomics, single-cell biology, proteomics, structural biology, and cheminformatics. These agents can spawn sub-agents and interact with specialist agents that users create. A dedicated reviewer agent checks citations and calculations, flagging and fixing errors on its own. Claude Science generates detailed scientific artifacts, including 3D protein structures, genome browser tracks, and chemical structures. Each output is paired with the exact code and environment used to produce it, alongside a plain-language explanation and the full message history. Anthropic’s incremental improvements to its agent tools, as seen with Claude Code, are built into this workflow.

Compute Management and Scalability

The platform manages computing resources across a user’s laptop, HPC cluster, or cloud GPU providers. When scaling up, Claude Science drafts a plan, asks for permission before tapping into new resources, and lets users review or revoke any decision before the job runs. It supports HPC clusters via SSH and on-demand compute through services like Modal. Meanwhile, Microsoft’s in-house models show that cost efficiency is possible with older hardware, matching frontier models at lower cost. Because its agents operate inside a running session that retains context in memory, large datasets only need to be loaded once. The platform runs on the user’s existing infrastructure, keeping sensitive data within trusted systems.

Domain-Ready on Day One

Claude Science comes pre-configured for biology and chemistry research, connecting to over 60 scientific databases, including UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, and GEO. It also integrates with NVIDIA’s BioNeMo Agent Toolkit to connect natively with life science models and libraries such as Evo 2, Boltz-2, and OpenFold3. Researchers can tailor the platform by saving pipelines as reusable skills or linking their preferred tools via connectors, settings that future sessions will automatically inherit.

Early Adopter Results

During beta testing, researchers used Claude Science for tasks ranging from single-cell RNA sequencing analysis to CRISPR screen design and protein structure prediction. Manifold Bio used the platform to nominate targets for tissue-targeting medicines. At the Allen Institute, Jérôme Lecoq built a multi-agent template to write long-form reviews, slashing a process that previously took up to two years to a fraction of the time. Stephen Francis, a UC San Francisco epidemiologist, used Claude Science to accelerate glioma molecular epidemiology studies, cutting analysis time by roughly 90 percent while producing results independently validated by his group. These real-world applications demonstrate the utility of agent-driven research, even as Chinese labs reach parity on benchmarks.

Availability and Grants

Claude Science is available in beta on macOS and Linux for Claude Pro, Max, Team, and Enterprise users. Team and Enterprise accounts require administrator approval. Anthropic is offering discounted Team plans for academic and nonprofit research labs. The company is also supporting up to 50 AI for Science projects with up to $30,000 in credits, with Modal providing up to $2,000 in compute for select projects. Applications are open through July 15, 2026, with projects running from September 1 to December 1, 2026. Anthropic’s support for science projects aligns with its broader policy strategy, doubling down on policy investments.