Anthropic announced the release of Claude Science, an application that brings together a researcher's scattered tools in one place and allows artificial intelligence agents to take over important parts of the work. This move marks the company's most significant entry into the laboratory environment. On June 30, 2026, Anthropic announced that Claude Science is now available in beta.
See also: Anthropic accuses Alibaba of the largest distillation campaign against Claude
The company describes it as an AI lab for scientists, integrating the databases, code tools, and computing resources that researchers manage every day.
An AI agent navigates between these resources. The application addresses a common problem faced by scientists who work across multiple databases, each with its own schema. Researchers frequently switch between platforms such as PubMed, Jupyter, R, and cluster terminals, often dealing with file formats that require custom workflows.
Claude Science simplifies these processes into a single environment, capable of analyzing the literature, conducting multi-step analyses, and refining figures and manuscripts until they are ready for publication. Importantly, Claude Science does not introduce a new model. It works using existing Claude models, including Opus 4.8, without special access. As TechCrunch notes, the emphasis is on improving workflow, not simply increasing model power.
At the core of Claude Science is a coordinating agent that uses over 60 selected skills and connections tailored for fields such as genomics, proteomics, structural biology and chemoinformatics. This agent can deploy additional agents, including specialized ones created by the user. A separate review agent verifies references and calculations, highlighting and correcting errors where necessary.
Anthropic emphasizes reproducibility, a critical issue in modern science. Each generated figure includes the exact code and environment used to create it, along with a simple explanation of its creation and a complete message history. Researchers can review results months later and trace their sources. They can also modify figures using plain language commands, asking the agent to adjust grid lines or change an axis to a logarithmic scale, with the agent automatically rewriting the corresponding code.
See also: Microsoft accuses Anthropic of presenting Claude as conscious
The review agent plays a critical role in ensuring accuracy, as AI models can sometimes construct references and schemas. The system examines results for unverifiable data and references that do not align with the code, with the goal of identifying errors before they are noticed by humans. Claude Science is designed to run on lab computers, running locally on macOS or Linux, or remotely via SSH or an HPC hub.
For large tasks, such as protein folding or genomics workflows, the agent formulates a plan and requests approval before accessing additional resources. It then submits tasks to the lab cluster or a Modal account for on-demand compute power, allowing for scaling from one GPU to hundreds. This architecture also addresses privacy concerns, as the application runs within the lab infrastructure, ensuring that large or sensitive datasets remain secure. Only the necessary context for each step is transmitted to Claude.
Researchers can also split a session to compare different approaches without losing the original data. The release is supported by a partnership with Nvidia. Claude Science uses BioNeMo Agent Toolkit to access life science models such as Evo 2, Boltz-2 , and OpenFold3 , and integrates with over 60 scientific databases, including UniProt, PDB , and ChEMBL.
Nvidia has invested in several areas of the AI industry, with life sciences being another area of focus. Early adopters of Claude Science include Manifold Bio, which designs targeted drugs. They used the app to identify targets for their latest experiments, taking into account factors such as surface expression, trafficking and safety. The company highlighted the app’s ability to perform tasks from start to finish, integrating the framework of previous programs.
See also: Claude Fable 5: Anthropic releases first public Mythos-Class model
Jérôme Lecoq, a neuroscientist at the Allen Institute, created a multi-agent model with about 20 customized skills for writing extensive reviews. The sub-agents read and analyze relevant literature, enhancing the research process.
