Anthropic Launches Claude Science: AI Workbench Tying 60+ Scientific Databases

Problem Addressed

Researchers in computational biology routinely juggle many tools—PubMed for literature, Jupyter for code, R for analysis, cluster terminals for computation, and separate software for protein structures—each with its own data formats and query methods. Claude Science aims to consolidate this fragmented workflow into a single environment.

Key Features

Claude Science does not introduce a new model; it uses the existing Claude models, including Opus 4.8, without dedicated biological training. A central AI Agent acts as a "project manager," connecting to more than 60 scientific databases covering genomics, single‑cell analysis, proteomics, structural biology, and cheminformatics. Users pose natural‑language questions, and the system dispatches specialized sub‑Agents to query and aggregate results.

Reproducibility . Every generated chart is accompanied by the full code, execution environment, a natural‑language description of the creation process, and the complete dialogue history, enabling exact reconstruction months later. Users can modify visual aspects (e.g., change the Y‑axis to a logarithmic scale) via natural language, and the Agent updates the underlying code automatically.

Local Compute . The platform can be installed on macOS or Linux and can connect via SSH to on‑premise HPC clusters, keeping sensitive data on local infrastructure. Only contextual information needed for each analysis step is sent to Claude. For heavy workloads, Claude Science can invoke a Modal account to scale out to hundreds of GPUs on demand.

Early User Feedback

Sean Whalen at Gladstone Institute built a genome browser from scratch in a few days. Stephen Francis at UCSF reported that Claude Science identified a year‑long viral contamination in their RNA‑seq data. Jérôme Lecoq at Allen Institute assembled a multi‑Agent literature‑review system that processed thousands of papers and generated a narrative review that previously took his team two years. MIT researcher Iain Cheeseman noted that the tool enabled him, a non‑computational biologist, to perform analyses that were previously impossible for him.

Competitors

OpenAI released GPT‑Rosalind in April, a model specifically trained for life‑science reasoning, with an upgrade in June. The two approaches differ: GPT‑Rosalind focuses on a domain‑specialized model, whereas Claude Science keeps the base model unchanged and concentrates on workflow integration, bundling databases, compute resources, and collaborative Agents into a platform.

GPT‑Rosalind is currently limited to U.S. enterprise customers under a research preview, while Claude Science offers a lower entry barrier: any Pro subscription (US$20/month) grants access.

This reflects Anthropic’s strategic shift from selling raw model capability to providing an industry‑specific operational layer, similar to how Claude Code became an operational layer for software development.

How to Use

Claude Science entered public beta on June 30 and is available for macOS and Linux. Access requires a Pro, Max, Team, or Enterprise subscription; Team and Enterprise plans need administrator enablement. Academic labs and non‑profit research groups can apply for discounted Team seats.

Anthropic will fund up to 50 Claude Science research projects, each up to US$30,000, with Modal contributing up to US$2,000 in compute credits. Applications close on July 15, decisions are announced by July 31, and funded projects run from September 1 to December 1.