Anthropic has unveiled Claude Science, a new application designed to streamline scientific research by integrating disparate tools into a unified AI-driven workbench. Announced on June 30, 2026, the beta release aims to address a persistent pain point for researchers: the fragmented workflow across databases, coding environments, and computing clusters. By centralizing these resources, Claude Science enables AI agents to handle complex, multi-step analyses, literature reviews, and manuscript preparation.
What is Claude Science?
Claude Science is not a new AI model but an application layer that runs Anthropic's existing Claude models, including Opus 4.8. It provides a graphical interface and agent-based system that can access over 60 curated scientific databases and tools, such as PubMed, UniProt, PDB, and ChEMBL, as well as programming environments like Jupyter and R. The core concept is to reduce the cognitive load on scientists by automating routine tasks like data retrieval, format conversion, and pipeline execution.
Key Features and Architecture
The platform is centered on a coordinating agent that can orchestrate a variety of specialist sub-agents. Users can create custom agents for specific tasks, such as protein folding or genomics pipelines. A particularly noteworthy component is the reviewer agent, which cross-checks citations, calculations, and code for errors. This agent is designed to catch AI hallucinations, such as fabricated references or illogical figures, before the human researcher reviews the output.
Reproducibility is a major focus: every generated figure includes the exact code and environment used to produce it, along with a plain-language explanation and the full message history. This allows researchers to revisit and modify results months later. Changes can be made via natural language commands—for example, asking the agent to adjust axis scales or remove gridlines—which the agent then implements by rewriting its own code.
Deployment and Privacy
Claude Science can run locally on macOS or Linux systems, or remotely over SSH or HPC clusters. This design ensures that sensitive datasets never leave the lab's infrastructure, addressing privacy concerns. Only minimal context is sent to Anthropic's servers. For large-scale computations, the platform can submit jobs to the lab's own cluster or to Modal for on-demand GPU resources, scaling from one to hundreds of GPUs.
Integration with Nvidia
A key partnership with Nvidia brings the BioNeMo Agent Toolkit to Claude Science, enabling access to specialized life-sciences models like Evo 2, Boltz-2, and OpenFold3. This integration allows researchers to run tasks such as protein structure prediction and molecular dynamics simulations directly within the workbench.
Early User Experiences
Anthropic highlighted three beta testers. Manifold Bio, a drug discovery company, used Claude Science to nominate targets for experiments by analyzing surface expression, trafficking, and safety data. Jérôme Lecoq, a neuroscientist at the Allen Institute, built a multi-agent template to write long-form literature reviews. His team previously spent up to two years on a single review; with Claude Science, they completed about ten reviews, each over 100 pages, in a fraction of the time. Stephen Francis at UCSF's Brain Tumor Center reported that a glioma analysis ran in roughly one-tenth of the usual time, and manual verification confirmed the results.
Broader Context and Implications
The launch is part of Anthropic's broader strategy to position Claude as a tool for genuine research, not just conversational AI. The company faces stiff competition from OpenAI, Google DeepMind, and others, but sees scientific workflows as a high-value market. The timing is notable given Anthropic's reported tensions with the US government over export controls on powerful AI models, which could complicate international collaboration.
Claude Science is available in beta for Pro, Max, Team, and Enterprise subscribers on macOS and Linux. Academic and nonprofit labs receive discounted pricing. Anthropic is also funding up to 50 research projects with $30,000 in compute credits each, with applications open until July 15, 2026. The scientific community will now test whether this AI workbench genuinely accelerates discovery or simply increases the volume of machine-generated output.
Source:TNW | Anthropic News