Anthropic Unveils Claude Science: An Integrated AI Workbench for Researchers

Overview

Claude Science, announced on 30 June 2026, is an AI‑driven workbench for scientific research. It integrates common research tools and software packages, enabling literature analysis, multi‑step experiments, manuscript drafting, and iterative figure refinement, with a complete, auditable provenance trail for every output.

Core Capabilities

Users interact with a general‑purpose coordination model that provides more than 60 predefined skills and connectors covering genomics, single‑cell analysis, proteomics, structural biology, and cheminformatics. The model can invoke other models, external platforms, and user‑created specialist models. An audit agent continuously checks citations, code consistency, and image provenance, flagging and correcting errors in real time.

Figure Generation and Editing

The system can generate scientific figures—including 3‑D protein structures, genome‑browser tracks, and chemical structures—and automatically attach the underlying code, environment details, and a concise creation description. Users edit figures with natural‑language commands (e.g., “remove grid lines” or “use logarithmic axis”), and the system updates the corresponding code automatically.

Compute Management

For large‑scale analyses such as protein folding or massive genomic datasets, Claude Science orchestrates compute jobs: creating tasks, dispatching them to clusters, monitoring completion, and retrieving results. Users may run tasks on their own HPC clusters or on‑demand resources via a Modal account, scaling from a single GPU to hundreds of GPUs. Session memory retains long‑term context, so large datasets are loaded only once and only task‑relevant information is transmitted to the model.

Workflow Examples

Researchers have used Claude Science for single‑cell RNA‑seq analysis, CRISPR screen design, protein structure prediction, and chemical informatics.

Manifold Bio employs Claude Science to design organ‑specific therapeutics, evaluate surface expression, delivery methods, safety, and to screen millions of candidate compounds using proprietary data.

The Allen Institute built a multi‑agent “computational review template” with roughly 20 specialized skills. Sub‑agents read thousands of papers, extract quantitative data into an evidence database, and assemble long‑form reviews by delegating sections to dedicated sub‑agents.

Availability

Claude Science is in testing on macOS and Linux and is offered under Pro, Max, Team, and Enterprise plans.

References

https://claude.com/product/claude-science

https://www.manifold.bio/

https://www.anthropic.com/news/claude-science-ai-workbench

Code example

来源：ScienceAI
本文
约2500字
，建议阅读
6
分钟
大模型在科学研究场景中的结构化协作能力与任务编排方式。