Anthropic Wants to Discover Its Own Drugs: Inside Claude Science

An AI lab decides to make its own medicine

Most of the frontier-model news cycle is about tokens per second and benchmark leaderboards. Anthropic just made a different kind of move. At a launch event aimed squarely at pharmaceutical executives, biotech founders, and bench scientists, the company unveiled Claude Science — a research environment it positions as the "Claude Code for science" — and, more provocatively, announced that it will use the tool to run its own internal drug discovery program targeting neglected diseases.

That second half is the story. Plenty of labs sell software to pharma. Very few AI companies have said, in effect, that they intend to try discovering drug candidates themselves. According to MIT Technology Review, which covered the launch, Anthropic's head of life sciences, Eric Kauderer-Abrams, framed the ambition in mission terms: "Our mission is to develop AI that serves humanity's long-term well-being, and we believe that by far the greatest opportunity to do that is in the life sciences." Whether that is genuine strategic conviction or a very well-produced narrative is the question worth sitting with.

What Claude Science actually is

Stripped of the framing, Claude Science is a research workbench built on Anthropic's existing models. Per Pharmaceutical Technology, it bundles more than 60 pre-configured functions spanning genomics, structural biology, proteomics, and cheminformatics, and wires in tools working scientists already live in — PubMed for literature, plus Jupyter and R for analysis. MIT Technology Review adds that it can run code on compute clusters and emphasizes reproducibility: the system is designed so researchers can trace a result back to its source rather than trust an opaque answer.

The demonstrated capabilities are concrete rather than sci-fi: designing CRISPR screens, analyzing single-cell RNA sequencing data, and rendering 3D protein structures. The pitch is that a scientist can hand Claude a high-level instruction and have it execute a multi-step workflow — pulling literature, running analysis, generating and refining figures — the way a software engineer delegates to Claude Code.

On availability, the sources are consistent but modest. Pharmaceutical Technology reports Claude Science is in beta, running locally on Linux or macOS or via a remote machine, and that it folds into existing Claude subscriptions rather than requiring a separate license. It builds on Claude for Life Sciences, which Anthropic debuted back in October 2025 — so this is an escalation of an existing bet, not a standing start.

The neglected-disease program

The internal drug program is what separates this from a product launch. Anthropic says it will use Claude Science to pursue candidates for neglected diseases — the conditions that traditional pharma has historically found commercially unattractive because the addressable market is small or poor.

The demonstrations gesture at what that looks like. MIT Technology Review reports the system showed identifying new drug candidates for phenylketonuria, a rare genetic disorder. MLQ's write-up describes a UCSF researcher using the tool to analyze 100 rare genetic diseases. These are demos, not clinical results — no candidate has been validated in a lab, let alone a patient, and the sources make no such claim. But the choice of neglected diseases is strategically shrewd: it is hard to accuse a company of profiteering when it points at conditions the market has abandoned.

There's also a stated feedback-loop logic. By doing drug discovery in-house, Anthropic argues it learns firsthand what tools and models pharma actually needs — which in turn should make Claude Science a better product for the paying biopharma customers it ultimately wants.

The customers and the credits

The commercial scaffolding is already visible. Named partners and customers across the sources include Novo Nordisk, the Allen Institute, and UCSF, and Pharmaceutical Technology notes Anthropic struck a deal with Bristol Myers Squibb in May 2026 to deploy Claude to "over 30,000 employees." That is a meaningful enterprise footprint for a product launching in beta.

To seed adoption among smaller labs, Anthropic is running a grant program: per MLQ, up to $30,000 in credits for roughly 50 research projects, with applications closing July 15 and recipients announced by July 31. It's a classic developer-relations playbook — buy your way into researchers' workflows early — transplanted into academic biology.

Hype versus what's real

Here's where a magazine has to draw a line. What's genuinely real: a shipping (beta) product, real integrations, named enterprise customers, and a credible existing base in Claude for Life Sciences. Anthropic is not vaporware-ing this.

What's aspirational: the drug discovery itself. Identifying a "candidate" is the very first step in a pipeline that runs through years of preclinical work, trials, and regulatory review — a pipeline where the overwhelming majority of candidates fail. None of the fetched sources claim Anthropic has produced a validated compound, and it would be a serious overreach to imply otherwise. The often-quoted line that AI compresses drug timelines has been made before by many companies; the industry's actual clinical track record for AI-originated drugs remains thin and contested.

One vivid data point is worth quoting precisely because of its ceiling. MIT Technology Review cites Harvard physicist Matthew Schwartz estimating that Anthropic's Opus 4.5 model is "about as capable of executing scientific projects as a second-year graduate student." That's genuinely impressive — and also a reminder of the level: a capable junior researcher who needs supervision, not an autonomous discoverer of cures. Read as praise, it's real; read as a promise of imminent breakthroughs, it's the hype talking.

Why this move matters anyway

Even discounting the drug claims, the strategic signal is significant. Anthropic is trying to move up the value chain — from selling tokens to owning scientific outcomes — and to differentiate on a vertical (life sciences) where reproducibility, tool integration, and trust matter more than raw speed. It also puts Anthropic explicitly in a three-way contest with Google and OpenAI, both of which have their own scientific-AI ambitions.

The risk is credibility. If Claude Science becomes an indispensable lab tool but the internal drug program quietly produces nothing, the narrative curdles. If, conversely, even one neglected-disease candidate advances, Anthropic will have a story no benchmark can match. The company has tied its most idealistic framing to its most uncertain bet.

The takeaway

Claude Science is a real, shipping research tool with real customers — that part isn't in doubt. The headline-grabbing piece, an AI lab discovering its own drugs for neglected diseases, is a serious ambition backed so far by demos, not medicines. The smart way to read it: Anthropic has built a credible scientific workbench and wrapped it in its most mission-driven story to date. Watch the product adoption, which is measurable now, and hold the drug-discovery promises to the only standard that counts in pharma — results that survive a trial, not a launch event.