Alongside Sonnet 5, Anthropic unveiled a second thing on June 30: Claude Science, an AI workbench for researchers. The idea is simple and hits a real pain point — scientific work is often tedious. Researchers juggle dozens of databases, each with its own schema, jump between PubMed, Jupyter, R, and a cluster terminal, and build bespoke pipelines for every file format.
Everything in one environment
Claude Science brings these fragmented tools into a single research environment. You analyze literature, run multi-step analyses, generate figures and manuscripts, and refine them iteratively until they’re ready to publish. The key trick: every output carries an auditable history — the exact code, the environment, and the full message history. That lets you validate and reproduce any result even months later.
Like a Jupyter Notebook, the app runs where you already work: locally on macOS or Linux, or over SSH on a remote machine or HPC login node. Sensitive datasets never have to leave your own systems — only the context needed for each step is sent to Claude.
Specialist agents, compute, and a reviewer
You work with a coordinating agent that has access to over 60 curated skills and connectors — pre-configured for genomics, single-cell, proteomics, structural biology, and cheminformatics. These agents can spin up others. For heavy compute, Claude Science plans it for you: it drafts a job, asks before reaching new resources, and scales from a single GPU to hundreds — via your own HPC cluster or a Modal account.
The clever part is the reviewer agent. It inspects the outputs, flags incorrect citations, untraceable numbers, and figures that don’t match their code — and self-corrects as it goes. Through NVIDIA’s BioNeMo Agent Toolkit, models like Evo 2, Boltz-2, and OpenFold3 are natively connected.
What researchers are already doing with it
Anthropic gives concrete examples. A neuroscientist at the Allen Institute built a multi-agent template of about 20 skills to write long-form reviews. What used to take up to two years now yields around ten reviews, many over 100 pages, with citations checked by reviewer agents. An epidemiologist at UCSF cut comprehensive germline workups on brain tumors to roughly one-tenth of the previous time — and validated the results independently.
Claude Science launches as a beta for Pro, Max, Team, and Enterprise on macOS and Linux. Research labs get discounted Team seats, and Anthropic is funding up to 50 projects with up to 30,000 dollars in credits each; applications run through July 15.
My take
This might be the most concrete “AI accelerates science” story I’ve seen in a while. Not “someday AI cures cancer,” but: a review agent catches wrong citations, and an analysis that used to take months runs in days. The auditable trail matters almost more than the speed — because without reproducibility, fast science is just faster wrong answers.
Sources: Anthropic: Claude Science, CNBC, Bloomberg