Anthropic introduces Claude Science, a workspace for AI laboratory development.

      Anthropic has introduced Claude Science, an application designed to consolidate a researcher’s various tools into a single platform, allowing AI agents to handle significant portions of their tasks. This marks the company's most substantial endeavor in the laboratory space to date.

      On June 30, 2026, Anthropic announced that Claude Science is available in beta. The company describes it as an AI workbench for scientists, integrating the databases, coding tools, and computational resources that researchers manage daily, with an AI agent facilitating transitions between them.

      This initiative addresses a genuine concern among scientists, who often navigate numerous databases, each with its unique schema. They frequently switch among platforms like PubMed, Jupyter, R, and cluster terminals while managing diverse file formats that necessitate tailored pipelines. Claude Science consolidates these processes into one unified environment, enabling literature analysis, conducting multistep analysis, and refining figures and manuscripts until they're publication-ready.

      However, it’s important to note that Claude Science is not a new model. It utilizes the existing Claude models available for sale, including Opus 4.8, without offering special access. As TechCrunch highlighted, the emphasis is on workflow rather than the raw power of the model.

      At the core of the application lies a coordinating agent, leveraging over 60 curated skills and connectors tailored for fields such as genomics, proteomics, structural biology, and cheminformatics. This agent can generate additional agents, including specialized ones created by the user. A distinct reviewer agent verifies citations and calculations, flagging and rectifying errors along the way.

      Anthropic is heavily invested in ensuring reproducibility, a major concern plaguing contemporary science. Each figure is accompanied by the precise code and environment that created it, along with a straightforward explanation of its creation process and full message history. Researchers can revisit results months later and trace the origins of any output. They also have the capability to modify figures using plain English; for instance, they can instruct the agent to remove gridlines or adjust an axis to a log scale, prompting it to automatically rewrite the corresponding code.

      The reviewer agent serves a crucial secondary function as well. AI models sometimes fabricate citations and numerical data. The system scrutinizes outputs for unverifiable figures and references that do not align with the code, aiming to correct errors before they are identified by a human reviewer.

      Claude Science is designed to operate on a lab's existing machines. It functions locally on macOS or Linux, or connects to a remote server via SSH or an HPC login node. Large tasks, such as protein folding or running a genomics pipeline, are managed by the agent, which outlines a plan and seeks approval before utilizing any new resources. The job is then submitted to the lab's cluster or to a Modal account for on-demand computation, capable of scaling from a single GPU to hundreds.

      This architecture also addresses privacy concerns. Since the application operates within the lab's infrastructure, large or sensitive datasets never leave the environment. Only the context necessary for each task is transmitted to Claude. Researchers can create separate sessions to evaluate two different approaches without losing the original data.

      The launch is supported by a collaboration with Nvidia. Claude Science employs the chipmaker’s BioNeMo Agent Toolkit to access life sciences models like Evo 2, Boltz-2, and OpenFold3, and it integrates over 60 scientific databases, including UniProt, PDB, and ChEMBL. Nvidia has invested in various sectors of the AI industry, with life sciences being another focus area.

      Early feedback from users of the beta version highlights its utility. Manifold Bio, which focuses on medication development targeting specific tissues, utilized Claude Science to identify targets for its latest experiments, considering surface expression, trafficking, and safety. The company noted that the app was advantageous because it could handle tasks comprehensively, incorporating the context of previous programs.

      Neuroscientist Jérôme Lecoq from the Allen Institute created a multi-agent template utilizing around 20 custom skills to compile long-form reviews. Sub-agents analyzed thousands of papers, extracted crucial findings, and compiled them in a database, subsequently drafting the review section by section. Lecoq mentioned that what once took his team as long as two years can now be completed in about a tenth of that time, allowing him to produce multiple reviews, some exceeding 100 pages.

      However, this increase in efficiency raises concerns. A tool that condenses a two-year review into a set of ten could accelerate genuine synthesis but might also lead to an inundation of machine-generated papers in an already burdened literature landscape. Anthropic’s solution to this issue includes the reviewer agent and human validations. Stephen Francis, an epidemiologist at the UCSF Brain Tumor Center, reported that his glioma analysis was completed in approximately one-tenth of the typical duration, and his team confirmed the results through manual checks.

      The launch aligns with a broader strategy. Anthropic positions Claude as a genuine research tool, not merely for conversation, aiming to validate this claim in