Anthropic has launched Claude Science, an app that pulls a researcher’s scattered tools into one place and lets AI agents run large parts of the work. It is the company’s biggest push yet into the lab.

Anthropic said on June 30, 2026 that Claude Science is now available in beta. The company calls it an AI workbench for scientists. It pulls together the databases, code tools, and compute that researchers juggle every day. An AI agent then moves between them.

The pitch targets a real complaint. Scientists work across dozens of databases, each with its own schema. They switch between PubMed, Jupyter, R, and a cluster terminal, and they wrangle file formats that need custom pipelines. Claude Science folds those steps into one environment. It can analyse the literature, run multistep analysis, and refine figures and manuscripts until they are ready to publish.

One thing it is not is a new model. Claude Science runs the same Claude models already on sale, including Opus 4.8, with no special access. As TechCrunch put it, the bet is on workflow rather than raw model power.

An agent that shows its work

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

At the centre sits a coordinating agent. It draws on more than 60 curated skills and connectors. These are set up for fields such as genomics, proteomics, structural biology, and cheminformatics. The agent can spin up other agents, including specialist ones built by the user. A separate reviewer agent checks citations and calculations, then flags and corrects errors as it goes.

Anthropic is leaning hard on reproducibility, the issue that haunts modern science. Every figure arrives with the exact code and environment that produced it. It also carries a plain-language note on how it was made, plus the full message history. A researcher can return months later and trace any result. They can also edit a figure in plain English. Ask the agent to drop gridlines or switch an axis to a log scale, and it rewrites its own code.

The reviewer agent matters for a second reason. AI models invent citations and numbers. The system inspects outputs for untraceable figures and references that do not match the code. It is meant to catch its own mistakes before a human does.

It runs where the data already lives

Claude Science is built to sit on a lab’s own machines. It works locally on macOS or Linux, or on a remote box over SSH or an HPC login node. Large jobs, such as folding a protein or running a genomics pipeline, fall to the agent. It drafts a plan and asks before reaching new resources. Then it submits the job to the lab’s own cluster, or to a Modal account for compute on demand. The work can scale from one GPU to hundreds.

That design also answers a privacy worry. Because the app runs on the lab’s infrastructure, large or sensitive datasets never have to leave it. Only the context needed for each step is sent to Claude. Researchers can fork a session to compare two approaches without losing the original.

The launch leans on a tie-up with Nvidia. Claude Science uses the chipmaker’s BioNeMo Agent Toolkit to reach life-sciences models such as Evo 2, Boltz-2, and OpenFold3. It also draws on more than 60 scientific databases, including UniProt, PDB, and ChEMBL. Nvidia has spread its money and tools across the AI industry, and life sciences is one more front.

What the early users say

Anthropic points to three beta users. Manifold Bio, which designs medicines that home in on specific tissues, used Claude Science to nominate targets for its latest experiments, weighing surface expression, trafficking, and safety. The firm said the draw was that the app could run the task end to end, with the context of past programmes built in.

Jérôme Lecoq, a neuroscientist at the Allen Institute, built a multi-agent template of about 20 custom skills to write long-form reviews. Sub-agents read thousands of papers, pulled the key findings, and stored them in a database, then drafted the review section by section. Lecoq said a single review used to take his team as long as two years. He now has about 10 of them, many running past 100 pages.

That number is also the catch. A tool that turns a two-year review into a batch of 10 could speed real synthesis. It could also flood an already strained literature with machine-made papers. Anthropic’s answer is the reviewer agent and human checks. Stephen Francis, an epidemiologist at the UCSF Brain Tumor Center, said his glioma analysis ran in about a tenth of the usual time, and that his group checked the results by hand and confirmed they held up.

A high-stakes bet on the lab

The launch fits a wider plan. Anthropic has framed Claude as a tool that can do real research, not just chat. Science is a market where that claim can be tested. It is also a commercial move. The company is racing to win paying customers ahead of a planned listing, and it has set out huge revenue targets to justify its spending.

The timing is awkward in one respect. Anthropic is in a tense standoff with Washington, after the US government moved to block foreign access to its most powerful models. A product built for open scientific collaboration lands in the middle of that fight.

Claude Science is in beta on macOS and Linux for Pro, Max, Team, and Enterprise plans, with discounted seats for academic and nonprofit labs. Anthropic will also fund up to 50 research projects with up to $30,000 in credits each. Applications are open until July 15, 2026. The bigger question is whether AI can truly speed discovery, or simply produce more of it. The labs now testing the app will give the first real answer.