Novo to partner with Anthropic on drug discovery efforts

Ronni Holmvig Strøm · 2026-09-20

Novo Nordisk and Anthropic announced a drug discovery collaboration on September 16. No therapeutic programs named, no benchmarks, no financial terms. What they committed to: Novo's

Novo Nordisk and Anthropic announced a drug discovery collaboration on September 16. No therapeutic programs named, no benchmarks, no financial terms. What they committed to: Novo's scientists and computational teams choose the discovery problems. Anthropic brings its frontier models and the Claude Science platform. Dario Amodei put a number on the ambition, saying AI carries "the potential to compress a century's worth of biological and medical breakthroughs into a decade."

Novo has run a version of this play before and won cleanly. In 2024 it built NovoScribe on Claude, Amazon Bedrock, and MongoDB, and cut the time to generate a Clinical Study Report from weeks to roughly ten minutes. A CSR reorganizes facts already sitting in the locked trial database into the shape a regulator will accept, and a reviewer can check the output against the source in an afternoon. Ground truth was in the same building.

Discovery pays out on a different clock.

A target hypothesis gets confirmed by an assay, then an animal model, then a human trial, and the honest verdict arrives years after the suggestion. We would call this the verification lag, and it is the engineering problem this partnership exists to attack.

Frontier deployments of the last couple of years have worked because the loop was short. Code compiles or it does not. A document matches its source or it does not. In a wet lab the loop closes on a timescale measured in fiscal years. So the load-bearing deliverable is an evaluation harness that separates a good hypothesis from a merely plausible one before the readout. That harness is worth more than any single molecule it clears, because every program after the first one inherits it.

Practitioners have been describing how to build it all year. Nicholas Larus-Stone, who runs AI at Benchling, gave the recipe in one line: "LLMs can work in the wet lab with the right harness and a human in the loop." Ryan Kosai of Potato filled in the missing half, arguing that models propose and read literature well, and that around them "you need deterministic tools, more of a traditional software approach to calculate and constrain things correctly."

Pharma has automated one workflow at a time, and value has stopped at the edge of each. This deal is scoped past that edge, which is why the clause that matters in the release is the one about problems "identified by scientists and computational teams at Novo."

Model access stopped being Novo's constraint a while ago. The company announced an OpenAI collaboration in April, named AWS a strategic partner in August, and has had Gefion supercomputer time through NVIDIA and the Danish Centre for AI Innovation since June 2025. Four partners in, the differentiator is internal: which questions get chosen, and whether the organization can recognize a real answer when a model hands one over. Taste in choosing the question is the scarce input, and it is the one Novo has spent four decades accumulating in metabolic disease.

The company is also spending it under pressure. Its US-listed shares closed at $41.71 on the day of the announcement, down from above $140 in 2024. Sales for 2026 are guided to fall by as much as 13%, and CagriSema missed non-inferiority against Lilly's tirzepatide. This week the firm shortened its name to simply Novo. A company in that position does not fund demonstrations. It funds instrumentation, because instrumentation is the part that compounds.

The wager pays the moment Novo can show, inside a single program's lifetime, that model-proposed hypotheses survive the assay at a higher rate than the ones its scientists would have picked alone. The ten-minute CSR was easy to believe because anyone could check it. The next win will be worth far more and look like far less, right up until a molecule no human proposed walks into a clinic.

Sources: Semafor, 2026, Novo Nordisk & Anthropic, 2026, R&D World, 2026.