Novo Nordisk spent the back half of the last decade as the most valuable company in Europe. It got there on the strength of one molecule, semaglutide, sold as Ozempic for diabetes and Wegovy for weight loss, and for a while the only real constraint on the business was how fast it could fill pens.
That period is over. The company has now told investors to expect 2026 sales to fall, and it has just signed a deal that says a lot about what it thinks comes next.
Novo Nordisk is putting Anthropic’s Claude into its research organization. The agreement, announced on September 16, covers drug discovery work and internal software development, and the centerpiece is a product most people outside biology have never heard of: Claude Science.
The short version
- Novo Nordisk will test Claude Science, Anthropic’s research workbench, on specific drug discovery workflows
- It will also use Anthropic’s frontier models for internal software development
- This is Novo’s third major AI partnership of 2026, after OpenAI in April and Amazon Web Services in August
- CEO Mike Doustdar wants Novo to be “the world’s most AI-driven healthcare company”
- The backdrop: Novo guided to a sales decline of 5% to 13% for 2026 and has lost most of its peak market value
- The thing AI has to beat: roughly nine out of ten drug candidates that reach human trials never get approved
- No financial terms were disclosed, and no molecule has been discovered yet
What Claude Science actually is
Anthropic released Claude Science in beta on June 30, and it is a different kind of product from the chatbot most people have used. It is a workbench, which in practice means it sits on top of the tools a working scientist already opens every day rather than asking them to paste things into a text box.
It ships pre-configured with more than 60 scientific databases and connectors covering genomics, single-cell work, proteomics, structural biology and cheminformatics. The architecture is agentic: a primary agent runs the project and either spawns specialized sub-agents to split up the work or hands tasks to custom agents a lab has built for itself.
The detail that matters most is the last one in the chain. Before anything reaches a draft, a separate reviewer agent audits the citations and the calculations. That exists because the single biggest objection scientists have to language models is that they invent references that look perfectly plausible, and a fabricated citation buried in a methods section is worse than no citation at all.
| What the partnership covers | What it does not cover |
|---|---|
| Testing Claude Science on selected R&D workflows | Any disclosed financial terms |
| Literature analysis and multi step research tasks | A named drug program or target |
| Producing figures and manuscript drafts that get iterated on | Anything that shortens a clinical trial |
| Frontier models for Novo’s internal software work | Regulatory submissions or approvals |
| Joint work on problems Novo’s computational teams pick | Exclusivity on either side |
Read that right hand column again, because it is the honest frame for every AI and pharma announcement you will see this year. What is being automated is the reading, the reasoning and the writing around an experiment. The experiment itself still has to happen, in a building, with people.
The wall every one of these deals runs into
Drug development is not slow because scientists read too slowly. It is slow because biology keeps saying no, and it keeps saying no late.
Look at where the red is. The failures that hurt are the late ones, and the late ones are expensive precisely because the company has already paid for everything upstream. A Phase III readout that misses is the single most costly event in the industry.
An AI workbench does very little about that. It can help a team read faster, spot a pattern across datasets that nobody had connected, and get from a question to a defensible figure in a morning instead of a fortnight. Those are real gains at the front of the funnel. They are not the same thing as knowing in advance that a molecule will work in humans.
Dario Amodei’s framing for this, quoted in the announcement, is that AI could “compress a century’s worth of biological and medical breakthroughs into a decade.” That is the maximalist version. The version Novo has actually signed up for is narrower and more testable: use the tool on some workflows, see what happens.
Why Novo is doing this now
The timing is not incidental, and you cannot read this deal properly without the financial picture behind it.
Eli Lilly’s tirzepatide brands, Mounjaro and Zepbound, overtook Ozempic and Wegovy as the fastest selling products in their categories. US pricing came down under the Most Favored Nation arrangement struck in late 2025. Semaglutide patents are starting to expire in some markets outside the United States. And CagriSema, the drug that was supposed to be the answer to all of that, has produced a run of results that investors have politely called mixed.
A company in that position has two ways to talk about the future. It can talk about the next molecule, which it does not yet have. Or it can talk about the machinery that finds molecules, which is what an AI partnership lets it do. Doustdar’s line about becoming “the world’s most AI-driven healthcare company” is a statement about R&D productivity, and productivity is the number Novo most needs to move.
Anthropic’s side of it
For Anthropic this is a reference customer in the most demanding enterprise category there is, and it arrives in the middle of a year where the company has been pushing Claude into places well outside the chat window. It has been building toward reading people’s linked bank accounts on the consumer side, and toward agentic research tooling on the professional side. Life sciences is the most credible of those bets, because the willingness to pay is enormous and the workflows are genuinely bottlenecked on reading and synthesis.
It also sits awkwardly alongside the rest of Anthropic’s year. This is the same company that, two weeks ago, banned several groups of working scientists from its platform over biological misuse concerns while admitting it could not prove intent in any of the cases. Both things are policy: open the model up to the biggest pharmaceutical customers you can find, and keep a hair trigger on everyone else doing biology. Those positions are reconcilable, but they are not comfortable next to each other.
How to read pharma AI announcements without getting sold
- Look for a named program. “Selected workflows” means a pilot. A target and an indication means commitment
- Look for money. Undisclosed terms usually means small, or mostly credits
- Check which end of the funnel it touches. Discovery tooling is cheap to improve and slow to show up in results
- Count the partners. A company with three AI partners in six months is buying optionality, not conviction
- Watch the R&D line, not the press release. If this works, it shows up as more shots on goal per dollar, three or four years out
The honest case for it anyway
It would be easy to file this as a struggling company buying a narrative, and that is part of it. But there is a real version of the argument underneath.
The stuff Claude Science automates is not trivial busywork. A research scientist spends an enormous share of their time reading papers, reconciling datasets that were never designed to talk to each other, rebuilding a figure for the fifth time, and writing up. None of that is the science. All of it is the tax on doing science. Cutting that tax meaningfully does not make biology more predictable, but it does mean a given team can run more ideas past the point where biology gets to render its verdict.
That is also the pattern showing up elsewhere in the industry. OpenAI spent the summer promoting what it called an automated research intern, and drew criticism for grading its own homework on the benchmark. The lesson there applies here. These tools are plausibly very useful, and the people selling them are currently the only people measuring them.
What to watch
- Whether a specific program gets named. The moment Novo attaches this to a disease area, the pilot has graduated
- Whether the OpenAI relationship survives. Novo signed with OpenAI in April for enterprise-wide work. Two frontier labs inside one R&D organization is unusual and probably temporary
- Whether Novo publishes anything. A paper with Claude Science in the methods section would be far more informative than any announcement
- What happens to headcount. Pharma has been cutting. “Productivity” is a word that cuts both ways
- Whether the reviewer agent holds up. The citation auditing layer is the load bearing claim. One high profile fabricated reference would set the whole category back
Novo Nordisk has not announced a drug. It has announced that it is going to try a tool, on some problems, with no money disclosed and no timeline given. In an industry where the gap between a promising compound and an approved medicine is measured in a decade and a billion dollars, that is a modest thing dressed in ambitious language.
The ambitious language is not wrong, though. If a workbench like this genuinely compresses the reading and reasoning work at the front of the pipeline, the effect will not be visible in a press release. It will be visible in how many candidates a company Novo’s size can put into the funnel four years from now, and whether any of them make it through the part where biology gets a vote.
