Anthropic Introduces Claude Science: The AI That Analyzes Data Without Leaving the Lab
Anthropic has launched the beta of Claude Science, a workspace for scientific research that the company places on the same level as Claude Code and Claude Cowork. This is not a new model: it runs on the same Claude models already available to everyone, including Opus 4.8, without special access or restrictions. Anthropic clarifies this in the announcement, stating that "it is not a new AI model nor a more capable model for biology."
Claude Science builds on Claude for Life Sciences, presented in October 2025, and represents its evolution into an autonomous product, the third that Anthropic defines as flagship alongside Claude Code and Claude Cowork. It has been available since June 30 as a desktop app for macOS and Linux for users of Pro, Max, Team, and Enterprise plans, and it integrates over 60 scientific databases with preconfigured connectors for genomics, single-cell, proteomics, structural biology, and cheminformatics.
A Multi-Agent Architecture with a Reviewer
The operation revolves around multiple agents. A generalist coordinating agent orchestrates specialized agents, which can also be customized by the user, while a separate reviewing agent checks citations, calculations, and figures, signaling and correcting errors. On the reproducibility front, every generated figure carries "the exact code and environment that produced it, a natural language description of how it was created, and the complete history of messages."
The workbench runs on the laboratory's infrastructure: a macOS or Linux laptop, a remote server via SSH, an HPC node, or compute on demand via Modal. Sensitive datasets never leave the laboratory systems. For life sciences models, Claude Science relies on NVIDIA's BioNeMo Agent Toolkit to connect natively to tools like Evo 2, Boltz-2, and OpenFold3.
Initial Use Cases and Market Challenge
Some of the reported results are remarkable. Jérôme Lecoq, a neuroscientist at the Allen Institute, constructed a multi-agent review template with about 20 customized skills: a task that previously could take up to two years now produces 10 reviews of over 100 pages with citations verified by the reviewing agents. Stephen Francis, an epidemiologist at the UCSF Brain Tumor Center, reduced analysis time to about one-tenth in a molecular epidemiology study of glioma, with results independently validated by his group.
The bet is on distribution rather than the model, and here three different paths open up in the same market. Anthropic bets on broad access via subscription; OpenAI, with GPT-Rosalind, has chosen a specialized model limited to qualifying U.S. enterprise clients; Google DeepMind continues with Gemini for Science and proprietary models like AlphaFold and AlphaGenome. Interestingly, the announcement of Claude Science comes shortly after the departure of John Jumper, co-recipient of the Nobel Prize in Chemistry for AlphaFold, from Google DeepMind to Anthropic, as we reported in previous weeks.
The product arrives alongside a support program: Anthropic will fund up to 50 research projects with credits of up to $30,000 each, to which Modal adds up to $2,000 in compute. Applications remain open until July 15, 2026, with notifications by July 31 and projects scheduled from September 1 to December 1. Meanwhile, the Team plan offers discounted spots for active labs in academic institutions and nonprofit research organizations.