Claude Science, an AI workbench for scientists
Anthropic ● Covered by 3 sources
Anthropic just launched Claude Science, an AI workbench that bundles lab tools, code, and compute into one app for researchers. It's already cut some analyses from years to weeks, which is the kind of claim scientists usually meet with heavy skepticism.
Based on reporting by Anthropic — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic has spent the past several months building out its life sciences push, and today that effort gets a proper product: Claude Science, a workbench app aimed at researchers who currently bounce between PubMed, Jupyter, R, cluster terminals, and dozens of siloed databases just to get through a single experiment. The pitch is consolidation. Instead of stitching together bespoke pipelines for every file format and schema, scientists get one environment, running locally on macOS or Linux or over SSH to a remote machine or HPC login node, where a generalist coordinating agent hands off work to more than 60 curated skills and connectors built for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics.
What makes this more than a chatbot wrapper is the traceability built into every output. When Claude Science generates a figure or a manuscript section, it keeps the exact code, environment, and full message history attached, so a result can be checked or reproduced months later. A separate reviewer agent runs alongside the main work, flagging bad citations, numbers that don't trace back to source data, and figures that drift from the code that made them. For big jobs, like folding a protein or running a genomics pipeline across a huge dataset, the app plans the compute job itself, asks permission before touching new resources, and submits it to whatever a lab already uses, an in-house HPC cluster or a Modal account, scaling from one GPU up to hundreds without the researcher babysitting the queue.
Anthropic also leaned on partnerships to make the tool useful on day one rather than requiring months of configuration. Through NVIDIA's BioNeMo Agent Toolkit, Claude Science connects to models like Evo 2, Boltz-2, and OpenFold3, and it can plug into a lab's own trusted pipelines and datasets as reusable skills that carry over into future sessions.
The early users are the more interesting part. Manifold Bio, which designs medicines that target specific tissues, used the app to rank candidate binders against criteria drawn from its own proprietary data, something the company says a generic coding assistant couldn't do end to end. At the Allen Institute, neuroscientist Jérôme Lecoq built a roughly 20-skill pipeline that reads thousands of papers, extracts claims and findings into an evidence database, and drafts long-form reviews section by section, with one agent writing and another checking accuracy. Work that used to take his team up to two years now produces reviews, several over 100 pages, in a fraction of the time. And at UCSF's Brain Tumor Center, epidemiologist Stephen Francis says the app cut germline variant workups on glioma susceptibility to roughly a tenth of their previous duration, results his lab then validated independently.
Claude Science is live in beta for Pro, Max, Team, and Enterprise users, with a discounted Team tier for academic and nonprofit labs. Anthropic is also funding up to 50 AI for Science projects with as much as $30,000 in credits each, plus up to $2,000 in Modal compute for select projects, with applications open through July 15, 2026 and projects running from September through December of that year.
My take — AI-written commentary, not fact-checked reporting
Cutting a two-year literature review down to weeks is the kind of number that should make people ask harder questions, not fewer, especially when the reviewer checking the work is itself an AI agent built by the same company selling the subscription. The reproducibility features and code traceability are the right instinct, but a workbench this deeply woven into how labs analyze proprietary data and publish results is also a serious dependency to hand over to one vendor. Useful tool, worth watching closely, not worth taking on faith just because it's fast.
Read more about this at: Anthropic
