Anthropic Launches Claude Science, an AI Workbench for the Lab

Anthropic has a template now. Build a purpose-made application on top of Claude, aim it at a profession that lives in a mess of fragmented tools, and let the model absorb the workflow rather than the task. It did this for software with Claude Code, and Claude Code became one of the fastest-growing products the company has ever shipped. On June 30, Anthropic pointed the same playbook at the laboratory. Claude Science, launched in beta, is an AI workbench for scientists — and it is the clearest signal yet that Anthropic intends to be more than a chatbot vendor.

The pitch is deceptively simple: give a working researcher one environment that does the whole arc of computational science, from pulling a dataset to producing a figure that survives peer review. In practice, that is an enormous amount of plumbing. A molecular biologist today bounces between database portals, Python notebooks, structure viewers, cluster job schedulers, and citation managers, losing hours to the seams between them. Claude Science's bet is that an agentic system can hold that entire stack in its head — and that the friction it removes is where the discovery time actually goes.

What it actually does

Four capabilities anchor the product, and each targets a specific point of pain.

The first is Rich Artifacts. Claude Science generates figures and full manuscripts backed by reproducible code, and it renders the objects scientists actually care about natively — 3D protein structures, genome tracks, chemical structures — inside the workbench rather than as flat images pasted from another app. The artifact is the code and the output together, which means a reviewer or a collaborator can rerun it.

The second is compute orchestration. The system scales work from a laptop up to HPC clusters or on-demand GPUs and manages job submission itself, so a researcher does not have to hand-configure a scheduler to run something heavy. Anthropic has wired in Modal for on-demand computing, which lets the workbench spin up GPU capacity as a job demands it instead of forcing every lab to own a cluster.

The third is domain depth. Claude Science ships pre-configured for genomics, single-cell analysis, proteomics, and cheminformatics, with what Anthropic describes as more than 60 curated skills and connectors. Those connectors reach the canonical databases of the field — UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO — so the model is querying primary sources rather than reciting half-remembered facts. Anthropic also integrated NVIDIA's BioNeMo Agent Toolkit, pulling in specialist models like Evo 2, Boltz-2, and OpenFold3 for sequence and structure prediction.

The fourth — and the one that matters most for whether scientists trust it — is a reviewer agent that checks citations and calculations, flagging and correcting errors before they propagate. Around it, Claude Science runs a multi-agent structure: generalist agents coordinate the workflow while researchers spin up their own specialist agents for a narrow task. The reviewer is Anthropic's acknowledgment that a confident, wrong scientific claim is far more dangerous than a confident, wrong line of code.

Claude Science runs on macOS and Linux and is available across Pro, Max, Team, and Enterprise plans, with discounted Team pricing for academic institutions and nonprofits — a pointed nod at the university labs that are the seedbed of the field but rarely the budget holders.

The real move: Anthropic in the lab, not just next to it

A workbench is a product. What makes this launch strategically loud is that Anthropic is not content to sell the shovels. The company has said it will use Claude Science to pursue its own research into drugs for rare and neglected diseases — the targets that traditional biopharma passes on because the economics do not clear. That is a lab company using its own model to run its own programs, and it changes the shape of the bet from tooling revenue to the possibility of owning discoveries outright.

The pieces have been assembling in plain sight. In April, Anthropic acquired the computational biology startup Coefficient Bio in a deal reported at roughly $400 million in stock. In June, it hired John Jumper, who led AlphaFold at Google DeepMind and shared the 2024 Nobel Prize in Chemistry, after he announced his departure from DeepMind on June 19. You do not buy a computational-biology team and hire an AlphaFold laureate to ship a text box. You do it to build a science company inside a model company.

The timing is not incidental either. Anthropic is widely reported to be moving toward an IPO, and Claude Science reads as an argument to prospective investors: the frontier lab that can turn agentic capability into pharma and enterprise revenue, not just per-token API calls. Drug discovery is one of the few markets large enough to matter at Anthropic's valuation, and one of the few where a genuinely better research loop translates directly into value.

The caveat that runs through all of it

The reviewer agent exists for a reason. Science's tolerance for fabrication is zero, and large language models still hallucinate citations, invent plausible-looking numbers, and paper over gaps in reasoning with fluent prose. A workbench that renders a beautiful protein structure and cites a paper that does not say what it claims is worse than no tool at all, because it launders a mistake through an authoritative interface. Anthropic clearly knows this — the verification layer is front and center — but the burden of proof sits with the labs now. The product will be judged not on the demo but on whether a result it produced holds up when someone reruns the code.

Claude Science is a beta, and betas overpromise. But the direction is unambiguous. Anthropic has decided that the most valuable thing a frontier model can do is not answer questions faster — it is to compress the distance between a hypothesis and a reproducible result. If that compression is real, the lab is exactly where it should show up first.