Anthropic Launches Claude Science, an AI Lab Workbench

Anthropic is taking Claude into the lab. On Tuesday, June 30, the company launched Claude Science, a research workbench that folds a scientist's scattered tooling — literature analysis, multi-step experiments, compute management, and publication-ready figures and manuscripts — into a single environment built for auditability and reproducibility. It is available in beta to Claude Pro, Max, Team and Enterprise users on macOS and Linux, Anthropic said.

The pitch is aimed squarely at why researchers distrust AI output. Claude Science natively renders rich scientific artifacts — 3D protein structures, genome-browser tracks, chemical structures and molecular visualizations — and Anthropic says every figure ships with the exact code, environment specifications, a plain-language description and the full message history behind it, so a result can be reproduced rather than taken on faith. A separate reviewer agent inspects the model's own outputs, flagging incorrect citations and mismatched figures and attempting self-corrections before a human ever sees them.

Crucially, the workbench meets labs where their data and compute already live. Claude Science can submit jobs to a researcher's existing HPC cluster over SSH, to a Modal account, or to a local machine, scaling from a single GPU to hundreds as a job demands — and Anthropic says sensitive datasets stay on that on-site infrastructure, with only the necessary context sent to Claude. It arrives with more than 60 pre-configured skills and connectors for genomics, single-cell analysis, proteomics, structural biology and cheminformatics, and can query specialized databases including UniProt, PDB, Ensembl, Reactome and ChEMBL. An integration with NVIDIA's BioNeMo Agent Toolkit wires in specialized models such as Evo 2, Boltz-2 and OpenFold3, while Modal is offering up to $2,000 in compute credits for select projects.

Anthropic backed the launch with early adopters. It says Manifold Bio used Claude Science to nominate therapeutic targets by assessing surface expression, trafficking and safety across millions of candidate binders; that the Allen Institute's Jérôme Lecoq built a multi-agent "computational review template" with more than 20 custom skills that now produces 100-plus-page reviews with verified citations, compressing what had been two-year timelines; and that the UCSF Brain Tumor Center's Stephen Francis accelerated germline variant analysis to roughly a tenth of its previous timeframe without sacrificing validity. To seed more of that, Anthropic opened an AI for Science program offering up to $30,000 in credits to as many as 50 projects, with applications due July 15 and a project window of September 1 to December 1.

The launch lands in a field that has learned to be skeptical of AI's scientific promises. Rivals are pushing model-first: OpenAI has widened access to GPT-Rosalind, its biology-tuned model, even as its own LifeSciBench showed the best model passing just 36% of life-science tasks, and a widely cited Nature study found human scientists still outperform top AI agents on complex research. Anthropic's answer is not a bigger brain but a better bench — one that keeps the data at home, shows its work, and checks its own citations.