Anthropic is going to make its own drugs. Read that again.

On the last day of June, at an event in San Francisco, Anthropic did two things at once. It launched Claude Science, an AI workbench built for drugmakers — a tool to help pharmaceutical scientists run the fragmented mess of modern drug discovery from one place. And it announced that it will start discovering drugs of its own, standing up an internal program aimed at ‘neglected’ diseases the big biopharma companies don’t consider worth the trouble.

Read those two sentences back to back and the shape of the thing appears. Anthropic is selling the shovels and staking a claim in the same gold rush. It will supply the reasoning tool to the entire industry — and compete with that industry using the same tool. That is not a scandal. It’s a strategy. But it is a perfect, on-the-record example of something we’ve been saying for months: when AI money moves into health, it almost never moves for one reason.

the official reason, and the quieter one

The stated logic is reasonable, even likeable. Anthropic’s life-sciences lead put it plainly: to build the right tools for the industry, we need to live it — to be ‘in the trenches trying to develop drugs’ alongside our customers, learning from tight feedback loops. As a public-benefit company, Anthropic says it can pick programs on patient benefit, ‘including work the commercial market overlooks.’ Neglected diseases are exactly the kind of thing the market ignores because the maths never works: too small a population, too little profit, too much risk.

Take all of that at face value — a lot of it is probably true. Then notice what the arrangement also does, whether or not anyone intended it. Every pharma company that adopts Claude Science pours its hardest problems, its proprietary context and its scientists’ judgement into a tool owned by a company that is now, itself, a drug developer. The feedback loop that makes the product better is the same loop that makes the in-house program smarter. ‘Neglected diseases’ is a genuinely good cause and a very good on-ramp: low competition, high goodwill, and a training ground for a capability that does not stay confined to neglected diseases.

“There is no rule that says a company gets one motive. The useful question is never ‘are they good or bad?’ — it’s ‘what does this arrangement reward, no matter what anyone intends?’”

vertical integration of intelligence

For a decade the story of AI in health was ‘software helps the experts.’ The lab kept its model at arm’s length; the model was a supplier. What Anthropic just described is different: the model maker moving down the stack to own the thing the model produces. The reasoning layer becomes the value-capture layer. Whoever owns the intelligence that finds the drug is now positioned to own the drug.

This is not unique to one company, and it isn’t limited to pharma. It’s the same move you can watch happening everywhere at once: the AI provider stops selling you a tool and starts occupying the outcome the tool was for. Alphabet, Apple and Amazon each spent years pushing into healthcare from different doors — devices, records, clinics, pharmacies. The frontier-model labs are now walking through a new one, and they’re bringing the reasoning engine with them. When the same handful of companies own the model, the workbench and the molecule, ‘competition’ and ‘infrastructure’ start to describe the same firms.

what this means for you, specifically

You are not going to out-compete a frontier lab on drug discovery, and you shouldn’t try. If Anthropic’s program produces a treatment for a disease pharma abandoned, that is a straightforwardly good outcome — take it, gratefully. The infrastructure layer doing infrastructure things is fine. The point isn’t to be afraid of it. The point is to see it clearly and to be deliberate about the one layer these companies are not coming to run: your own understanding of your own body.

Here’s the practical translation. The tools you use to reason about your health — the general-purpose chat models you paste your labs and your training data into — are built by companies that increasingly have their own pipelines, their own partnerships, their own outcomes to capture. That doesn’t make the tools untrustworthy. It makes them interested parties. Use them the way you’d use any capable advisor whose employer has a side business: for what they’re good at, with your eyes open about whose incentives sit behind the interface.

1. Separate the layers on purpose. The infrastructure layer — drug discovery, protein design, the models themselves — is out of your hands and mostly fine there. The personal layer — your data, your weekly decisions, the questions you bring to a clinician — is yours. Don’t let a headline about the first convince you to hand over the second.
2. Ask who owns the output. When any AI product offers to ‘handle’ part of your health, ask what the company that built it also sells. A tool is not disqualified by having a business model — but you should know what it is.
3. Keep your data somewhere you can read it yourself. The single best hedge against every incentive you can’t see is being literate enough about your own numbers that no interface is the only thing standing between you and understanding.
4. Use AI as a thinking partner, not an authority. Ground it in your own records, use it to understand and to prepare better questions, and leave diagnosis to the qualified human. That division of labour survives every reorganisation of the industry above it.

the pattern, one more time

None of this requires believing Anthropic is acting in bad faith. It almost certainly isn’t. The discipline is simpler and more durable than suspicion: assume good products can have layered motives, watch where the value is being captured, and refuse to outsource the one part of your health that was always going to be yours to keep. The labs are moving down the stack to own the molecule. Let them. Just make sure you still own the judgement.

AI arriving in health from the top — through capital, through workbenches, and now through the molecules themselves — doesn’t shrink the value of understanding your own. It raises it. The more of the stack a few companies own, the scarcer, and the more valuable, the part they can’t: knowing what to ask, and when to trust the answer.