If you picture a research biologist at work, you probably imagine someone in a crisp white coat holding a test tube up to the fluorescent lights. The reality, however, is often a lot more mundane—and frankly, a lot more frustrating. Modern scientific discovery is less about sudden “Eureka!” moments in the lab and more about staring at a dual-monitor setup, wrangling dozens of massive databases, fighting with bespoke data pipelines, and constantly alt-tabbing between PubMed, Jupyter notebooks, and a cluster terminal just to get a single analysis to run.

Anthropic wants to fix that. The company has officially rolled out its most ambitious vertical product yet: Claude Science, an AI workbench built entirely around the messy, complex, and computationally heavy workflows of actual researchers.

Instead of just offering a smarter chatbot, Anthropic is pitching Claude Science as a centralized, generalist coordinating agent. The idea is to take all those fragmented tools that slow researchers down and bring them into a single environment. Available now in beta for macOS and Linux, the app comes pre-configured with over 60 curated skills and connectors geared toward fields like genomics, proteomics, and cheminformatics.

What makes this launch particularly interesting is how deliberately it targets the specific pain points of the scientific community. In science, if you can’t reproduce a result, it essentially didn’t happen. AI models have historically struggled with this, often acting as black boxes that spit out answers without showing their work. Claude Science is designed to produce rich, natively rendered scientific artifacts—like 3D protein structures and genome browser tracks—alongside an auditable history. When it generates a figure, it provides the exact code and environment used to make it, a plain-language explanation of its process, and the full message history. It even employs a dedicated “reviewer agent” that acts as a built-in fact-checker, automatically flagging incorrect citations, untraceable numbers, and calculations that don’t align with their underlying code.

Then there is the issue of data privacy and raw computing power. Researchers routinely handle massive, highly sensitive datasets, like patient genomes, that simply cannot be uploaded to a public cloud. Claude Science bypasses this by running locally on a lab’s own infrastructure—whether that’s a laptop or an HPC login node. The datasets stay exactly where they are, and only the necessary context for each analytical step is sent to Claude. If a job requires serious heavy lifting, like running a massive genomics pipeline, the AI drafts a compute plan and, with the user’s permission, seamlessly routes the job to the lab’s existing HPC cluster or scales up on-demand GPUs via integrations like Modal.

The early results from the beta phase suggest this isn’t just a quality-of-life upgrade; it’s a massive accelerator. Take Jérôme Lecoq, a neuroscientist at the Allen Institute. Writing a comprehensive scientific review traditionally took his team up to two years. Using Claude Science, they built a custom multi-agent workflow where sub-agents dig through thousands of papers to extract key claims, while reviewer agents rigorously check the citations for accuracy. Today, his team has churned out roughly ten reviews—many exceeding 100 pages—in a fraction of the time it used to take.

Similarly, Stephen Francis, an epidemiologist at the UCSF Brain Tumor Center, used the workbench to map out the genetic susceptibilities of glioma, a type of brain tumor. By automating the comprehensive germline workups, his lab completed their analysis in roughly one-tenth the usual time, with independent validations confirming the AI’s results were both rapid and robust.

Anthropic isn’t building this ecosystem in a vacuum. Acknowledging that scientists already have pipelines and open-source models they rely on, Claude Science integrates with NVIDIA’s BioNeMo Agent Toolkit. This gives researchers native access to industry-standard models like Evo 2 and OpenFold3 directly within the workbench.

The launch of Claude Science signals a fascinating shift in the AI industry. We are moving past the era of generic assistants trying to be everything to everyone, and entering a phase of highly specialized, domain-specific tools. To kickstart adoption, Anthropic is offering up to $30,000 in credits to 50 early research projects, with applications open through mid-July. If this workbench works as well at scale as it has in beta, it might just mean researchers can finally stop fighting with their file formats and get back to actually doing science.