Anthropic — the $183B AI lab behind Claude — has told the Financial Times it will start developing its own drugs. Not tools for pharma. Not copilots for researchers. Actual therapeutics, discovered and progressed in-house, with a team lifted from the pharma industry.
Read that again. The company selling the model is now also selling the molecule the model designs.
We have been saying for a year that AI investment into health and wellness rarely arrives clean. The pitch is always the same — faster trials, cheaper discovery, personalised medicine. The quiet part is that whoever owns the model ends up owning the downstream asset, because the marginal cost of running one more experiment collapses toward zero and the marginal value of a patented compound does not.
The company selling the model is now also selling the molecule the model designs. That is not a tool business. That is a pharma business wearing a hoodie.
Why this was inevitable
Foundation model economics are brutal. Training runs cost billions. Inference margins get compressed every quarter by the next open-weights release. The only way to justify the capex is to own something the model output is worth more than the tokens it took to produce it.
Drugs are that something. A single approved therapeutic can gross more than an entire year of API revenue. Isomorphic Labs — Alphabet's DeepMind spin-out — figured this out in 2021. Recursion, Insitro, Xaira and now Anthropic are all running the same play: use the model as a moat, keep the IP, license or commercialise the compound.
The Financial Times piece notes Anthropic has hired Anthropic's first Chief Science Officer from the pharma side and is building wet-lab partnerships. That is not a research collaboration. That is a pipeline.
What it means for the person actually taking the pill
Three things get worse before they get better.
Trial design bias. When the same lab designs the discovery model, picks the target, runs the in-silico screen and writes the trial protocol, the entire evidence chain flows through one set of priors. Regulators are not resourced to audit that. The FDA has one full-time AI reviewer for every several hundred submissions. The evidence looks rigorous because the maths is rigorous. The question of what was never tested is much harder to see.
Price anchoring on "AI-discovered". Expect a premium tier. "Designed by a frontier model" will get quoted the way "biologic" got quoted in the 2010s — as a reason a $6 pill costs $600. The marginal cost argument (models are cheap) will not pass through to the patient. It never does.
Consolidation of what counts as health. If the labs that own the LLMs also own the drugs, the LLMs will — subtly, then obviously — recommend the drugs. Not through crude product placement. Through what gets surfaced when a user types "I can't sleep" or "my joints hurt" into a chat interface that half the working population now uses as a first-line GP.
The FDA has one full-time AI reviewer for every several hundred submissions. The evidence looks rigorous because the maths is rigorous.
The wellness angle everyone will miss
The obvious story is pharma. The story we care about is what happens one ring out.
If a frontier lab can design a molecule, it can design a supplement stack. It can design a peptide protocol. It can design a "recovery formulation" that skirts the drug/food line the way NMN and rapamycin already do. The regulatory bar is a tenth of what a real drug faces. The margins are similar. The distribution — direct to consumer via the same chat interface — is already built.
Watch for the first AI lab to quietly launch a longevity brand. Not a partnership. A wholly-owned SKU. It will be dressed up as a research programme, priced like a luxury nootropic, and sold through an app most of its buyers already have open.
What sensible people should do
Three things.
One — treat any AI-mediated health recommendation as a paid placement until proven otherwise. Ask which lab owns the model. Ask whether that lab has a therapeutic pipeline. If yes, discount the recommendation.
Two — favour practitioners over pipelines. A coach, a physio, a nutritionist, a GP with a caseload has an incentive structure you can actually read. A model whose parent company sells the thing it just recommended does not.
Three — keep your own data. The clearest defence against a model that is quietly optimising for a downstream sale is a second opinion built on numbers only you hold. Sleep, HRV, bloods, symptoms, what actually worked last time. Bring that to any conversation — human or machine.
Where this ends
Every prior wave — search, social, mobile — collapsed the distance between the tool and the thing it was recommending. AI in health will do the same, faster, with higher stakes. Anthropic is not the villain of that story. They are the first mover confident enough to say the quiet part out loud. Everyone else in the frontier lab club is doing the same maths this quarter.
The question for the next decade is not whether AI will design our drugs. It is whether the same company can be allowed to design the drug, recommend the drug, dispense the drug and grade its own homework. The answer, historically, has always been no. We will find out whether that principle survives contact with a $183B valuation.
Sources
- Financial Times, "Anthropic to develop its own drugs" — ft.com
- Isomorphic Labs pipeline disclosures, 2024–2025
- FDA CDER AI/ML review capacity briefings, 2025
