Anthropic just dropped a beta version of Claude Science, an AI-powered workstation for scientists. The system runs locally, over SSH, or on HPC nodes, pulling together scattered databases, Jupyter notebooks, R environments, and compute terminals into one unified interface.
Its standout feature tackles a huge pain point in research: reproducibility. Every chart, 3D protein structure, or molecular model comes with the exact code, runtime environment, and full conversation history that produced it. Researchers can tweak visuals using plain language and trace back every data point.
On the compute side, Claude Science automatically submits and manages jobs on your lab's HPC cluster or Modal cloud, scaling from a single GPU to hundreds. Inside, there's a coordinator agent that taps into NVIDIA's BioNeMo toolkit and over 60 preconfigured scientific skills for analyzing gene and protein data — while a separate reviewer agent double-checks calculations and paper citations, catching errors on the fly.
Early test results are impressive. At UCSF's Brain Tumor Center, researchers cut analysis time for glioma studies to one-tenth of what it was (a 10x speedup). The Allen Institute used the multi-agent workflow to build a computational review template that shoved a 100-page academic review — which used to take two years — forward in a fraction of the time, with the reviewer agent verifying every citation's fidelity.
Claude Science is available now to Claude Pro, Max, Team, and Enterprise subscribers. Anthropic is also funding 50 research projects with up to $30,000 each, and its partner Modal is chipping in up to $2,000 in compute credits per project.