Anthropic launched Claude Science on June 30 as a beta workbench for researchers. The product brings Claude into a research environment that can connect to scientific tools, produce auditable artifacts, and run locally, on remote machines over SSH, or through an HPC login node.
The important part is the container around the model. Anthropic is not only saying Claude can help with scientific questions. It is trying to give researchers a place where literature review, data analysis, code execution, artifact creation, and review happen in one traceable workflow.
That is the right shape for serious scientific AI. A good answer is not enough if nobody can inspect where it came from.
The workbench is the claim
Anthropic says Claude Science gives users a generalist coordinating agent with access to more than 60 curated skills and connectors for areas such as genomics, single-cell work, proteomics, structural biology, and cheminformatics. The system can also work with specialist agents created by users.
That matters because research work is fragmented by default. A scientist may move between PubMed, Jupyter, R, a cluster terminal, domain databases, file viewers, manuscript drafts, and figures. A chat window can help with some of that work, but it does not naturally preserve the workflow.
Claude Science is aimed at that gap. Anthropic says outputs carry an auditable history so researchers can validate and reproduce results. The product also includes a reviewer agent that checks citations and calculations and can flag or correct errors.
The reviewer-agent detail is not a small feature. In science, the artifact is often less important than the review trail. If a model helps produce an analysis, the researcher still needs to know which source, tool, dataset, assumption, and calculation led to it.
Access is still narrow
Claude Science is available in beta for Claude Pro, Max, Team, and Enterprise users. Anthropic is also running a separate application program for up to 50 AI-for-science projects. Applications are open through July 15, 2026.
Selected projects can receive up to $30,000 in Claude credits. Anthropic also says select projects can receive up to $2,000 in Modal credits for serverless GPU compute. That turns the announcement into a product beta and a small funding mechanism at the same time.
The scope is useful, but still bounded. This is not a public claim that Claude can replace scientific review. It is a research-workflow product with compute access, tools, and review aids. Anthropic’s customer examples, including work with Novo Nordisk, the Novo Nordisk Foundation, the Broad Institute, and Arc Institute, should be treated as company-reported evidence unless outside evaluations follow.
Scientific AI needs provenance
Claude Science arrives one day after OpenAI published GeneBench-Pro, a benchmark aimed at judgment-heavy computational biology. The two stories point in the same direction from different angles. OpenAI is trying to measure whether models can handle scientific ambiguity. Anthropic is trying to give researchers an environment where model-supported scientific work can be inspected.
Both are more useful than the broad claim that AI can do science. The real question is narrower: can a model help scientists move faster while leaving enough evidence for another expert to check the work?
