Anthropic built a biology lab and will not say what it is for. Why?

TL;DR: Anthropic has built a wet lab in the San Francisco Bay Area and is teaching Claude to drive the robots in it. A spokesperson said it is not for drug discovery specifically, then declined to explain further.

My read: the lab is a data factory. AI has already scraped the whole internet, and it needs new data points. So some buy video of people doing their home routines, and others set up labs to run experiments.

Anthropic will confirm the lab exists. It will not say how big it is, how many people work there, when it opened, or what biosafety level it runs at. Biosafety level is the number that tells you what the room is allowed to handle.

Reuters broke the story on 18 September. A wet lab: a room for real experiments rather than simulations, somewhere in the San Francisco Bay Area, with Claude being taught to direct the robots running them. Eric Kauderer-Abrams, who leads Anthropic’s life sciences work, confirmed it. A spokesperson then added that the lab is not for drug discovery specifically, and declined to elaborate.

Language models were built on text that was already written down and already public. Public data has no moat, because it is public, and once the crawl is done every well-funded rival holds the same corpus.

You cannot look up how a given protein behaves in a given tube. Someone has to go and find out, and the measurement then exists in exactly one place, which is the lab that ran it. A rival who wants it has to buy the reagents and spend the instrument hours. Robotics hit this wall first, which is why Figure pays people to film their own chores.

A cookbook records the dishes that worked. The burnt ones go in the bin, so the book quietly teaches you that everything succeeds. Scientific literature has the same bias: results get published when they are worth publishing. A robot lab keeps the burnt ones. Every failed run is a labelled data point, and a model trained on a hundred thousand failures knows where the edges are in a way a model trained on papers cannot.

Medra runs what it calls the largest autonomous lab in the United States, and sells the output as a data foundry for foundation model teams. By its own estimate, after two decades of lab automation only about 5% of lab instruments are automated.

So when a spokesperson rules out drug discovery “specifically” and stops there, my read is boring: the lab is for training models on reality.

Anthropic is not running clinical trials “for now”, and Kauderer-Abrams says the line was drawn “due partly to competition concerns”. A competitive boundary is a commercial decision, and commercial decisions get revisited when the commercials change.

Amazon is the obvious precedent. The Wall Street Journal documented employees using third-party seller data to build competing own-brand products. Amazon denied it, and the EU case closed. The denials do not change the shape of it. A platform sees demand earlier than anyone selling on it, and eventually acts on what it sees.

The robot arms that make the data also remove the barrier that hands used to be.

Biosecurity has an old and reassuring argument. Information has sat in libraries for decades. The barrier was the skill: years at a bench before an experiment works at all.

Pipetting is the standard example. Abi Olvera of the Golden Gate Institute for AI quotes a working scientist calling it “mystical, if not lightly feared”. Students sometimes need more than fifty attempts before it works. A written protocol says “pipette gently”. How gently is thumb pressure, learnt the way a guitarist learns string pressure, and it changes with every buffer.

Robots do not have thumbs, and for years that was the whole reassurance. In August Anthropic published the Model Hardware Standard. One interface drives every piece of lab equipment, from pipetting robots to centrifuges, and integration work drops from weeks or months to hours or minutes. It shipped with device-level safety limits, more care than most standards arrive with.

Two weeks later Anthropic published a threat intelligence report on five blocked cases of people using Claude for biology research with weapons-relevant risk, one of them a grant proposal for gain-of-function work, meaning research that deliberately makes a pathogen more transmissible. Quote from the report: “Older models were well below the threshold where they could meaningfully assist in bioweapons development. This is no longer a certainty with newer models.” It is the bluntest version of that admission so far, though not the first. Anthropic shipped Claude Opus 4 under heightened protections in 2025 because it could not rule the same risk out.

The safety question has moved from what a model says in a chat window to which instruments it is allowed to touch.

Pharma avoids rare diseases for a clean economic reason. Development costs nearly as much as it does for a common disease, and the patient count is tiny. There are more than 10,000 rare diseases and fewer than one in ten has an approved treatment, with the neglect concentrated in the smallest populations. Cut the cost of discovery by a large factor and more of them become worth doing. Anthropic named that prize itself in June, saying it would do preclinical work in areas the industry does not find financially attractive.

Preclinical work is about a third of the cash cost of developing a drug. Human trials are the rest of the money and most of the years, and AI does not shorten them.

One of the Reuters sources gave the motive plainly. Anthropic wants to get there “before staff and the public lose faith that AI justifies erasing jobs and possibly human life”.

“We replaced your accountants” does not sell to the public. “We cured a disease that had no cure” does. The first AI company with a real drug will not be buying market share. It will be buying permission. And that comes in handy if you’re planning a Nasdaq listing at double your last private valuation.

The lab and the cure are one project. Data is what the machine needs, a cure is what the public will accept in exchange. Anthropic has planted itself exactly at the overlap of health and robotics - sectors where AI is not moving fast, held back by either regulation or complexity. And if there is not much political readiness or willingness to regulate AI properly worldwide, hopefully we’ll get rare diseases cured faster than some scary scenario evolves from the other end of it.