Anthropic launched Claude Science — and quietly became a drug company

The short version

Anthropic unveiled Claude Science, a research workbench built for scientists and the biopharma industry, and admitted it’s now running its own internal drug-discovery programs. That second part matters more than the product launch. Anthropic is no longer just selling AI to scientists. It’s using its own AI to try to find drugs, which puts it in the same room as the customers it’s pitching.

What did Anthropic actually announce?

Anthropic announced Claude Science, a research-focused workbench aimed at scientists and pharmaceutical companies, and at the same time revealed it has spun up internal drug-discovery programs of its own. The company is framing Claude Science as a new flagship product and a serious enterprise-revenue play, timed ahead of a rumored IPO. Within hours the news pulled coverage from STAT, CNBC, Endpoints and MIT Tech Review, which tells you the biopharma press took it seriously and not as another chatbot demo.

The workbench part is the easy story to tell. Researchers get a Claude-powered environment built around scientific work instead of general chat: reading literature, reasoning over experimental data, helping design and interpret studies. That’s a natural extension of what Anthropic already does, and pharma companies have been quietly wiring large language models into their pipelines for a while now. A polished, science-specific product with Anthropic’s name on it is an obvious next step.

The part that made me stop scrolling is the second announcement. Anthropic is starting its own drug programs. Not selling shovels to the miners. Mining.

Why does Anthropic running its own drug programs matter?

Anthropic running its own drug-discovery programs matters because it changes what kind of company Anthropic is. Selling an AI workbench to Pfizer or a biotech startup is a tools business. Running your own drug pipeline is a drug business. Those are two very different bets with very different timelines, risk profiles, and, frankly, conflicts of interest.

Here’s the honest read. When a company sells a tool to an industry and also competes in that industry, the customers start doing math. If I’m a mid-size biotech and I hand my proprietary research context to Claude Science, and Anthropic is simultaneously chasing its own targets, I’m going to ask hard questions about data separation, about what the model learns, about whether I’m training a competitor. Anthropic will have answers and contractual walls, and they’ll probably be real. But the question doesn’t go away just because the answer is good.

The generous interpretation is that Anthropic is running drug programs as proof. If your AI can actually help discover a drug, the best sales pitch is a drug. It’s the same move OpenAI makes when it builds flashy demos to sell the underlying capability. Dario Amodei has talked for a while about compressing decades of biomedical progress into years, and you can’t credibly claim that from the sidelines. You have to get in the arena and either produce a molecule or explain why you didn’t.

The less generous interpretation is that the tools business alone doesn’t justify the valuation Anthropic wants at IPO, and owning drug assets is a way to point at enormous future upside. Drug programs are lottery tickets, but they’re lottery tickets investors know how to dream about.

Is this actually good for science, or just good for Anthropic?

For working scientists, a genuinely capable research workbench could be good for science regardless of Anthropic’s motives. The bottleneck in a lot of research isn’t ideas, it’s the slow grind of reading everything, cross-referencing it, and spotting the connection three papers apart that nobody noticed. A model that’s decent at that gives a small lab some of the reach of a big one. That’s a real democratizing effect, and I don’t want to be cynical about it.

But good for science and good for Anthropic aren’t the same axis, and Claude Science sits at the intersection where they can pull apart. If the workbench becomes the default place research happens, Anthropic ends up with a view into what the entire field is working on. Even with perfect data hygiene on any single account, the aggregate signal of “what are thousands of scientists asking about right now” is enormously valuable to a company that also runs its own programs. That’s not a hypothetical abuse. That’s just the structural position the product creates.

My honest take is that the workbench is probably a net positive for individual researchers today, and the drug-company move is the thing to watch skeptically over the next few years. One helps you tomorrow. The other reshapes who holds power in biomedical research.

What does this mean for the AI-in-healthcare race?

The AI-in-healthcare race just got a new kind of competitor. Until now the landscape had roughly three camps: AI-native drug-discovery startups like Isomorphic Labs and Recursion, traditional pharma buying or building AI internally, and foundation-model labs selling capability. Anthropic just stepped into all three at once. It’s a foundation-model lab selling a tool, running its own pipeline like a startup, and pitching enterprise pharma like a vendor.

That’s an unusual amount of surface area, and it invites an obvious question I keep coming back to: can one company be the best AI provider to an industry and one of the best players in that industry at the same time? Amazon manages to sell AWS to companies it competes with in retail, so it’s not impossible. But drug discovery is far more sensitive about proprietary data than a warehouse is, and trust is the whole product.

For the rest of us, the practical implication is that the term “AI company” is getting slippery. Anthropic making its own drugs is a sign that the biggest labs won’t stay content selling access. They’ll integrate downstream into whatever industry has the fattest margins, and healthcare is near the top of that list. Expect more of this, not less, from every lab with the compute to try it.

What should you actually do about this?

If you’re a scientist or run a lab, treat Claude Science as a tool worth testing and worth reading the data terms on twice. The productivity upside is real, and so is the question of what you’re feeding into a model owned by a company with its own drug ambitions. Both things are true. Pilot it on non-sensitive work first and see if the reasoning holds up on problems where you already know the answer.

If you’re just watching the AI industry, the signal here isn’t the product. It’s the strategy. When Anthropic tells the story around its rumored IPO, whether investors reward it for the tools revenue or for the drug pipeline will tell you a lot about where the money thinks the value in AI actually lives. My bet is they’ll want to hear about the drugs, and that tells you where this is all heading. You can read the framing yourself in the original report from MIT Technology Review, which lays out both halves of the announcement.

FAQ

What is Claude Science? Claude Science is Anthropic’s research-focused AI workbench built for scientists and the biopharma industry, designed around scientific work like literature review, data reasoning, and study design rather than general chat.

Is Anthropic really making its own drugs? Yes. Alongside the Claude Science launch, Anthropic disclosed it is running internal drug-discovery programs of its own, which puts it in competition with some of the pharma customers it’s selling to.

Why is this happening now? Anthropic is positioning Claude Science as a flagship enterprise-revenue product ahead of a rumored IPO, and owning drug programs gives investors a large future-upside story to point at.

Should scientists use Claude Science? It’s worth testing for the productivity gains, but read the data terms carefully. Anthropic running its own drug programs makes the question of how your research data is handled more important than usual.

Does this create a conflict of interest? Structurally, yes. Selling an AI tool to an industry while competing in that industry raises real questions about data separation and trust, even if Anthropic maintains strict contractual and technical walls.