Big Pharma's AI Arms Race: Why Novo Nordisk Just Bet on Claude

On 16 September 2026, Novo Nordisk announced that it would use Anthropic's Claude models to help find new medicines (BioSpace). Read on its own, that sounds like a routine corporate press release. But Novo signed a separate, wide-ranging AI partnership with OpenAI only five months earlier, in April 2026 (CNBC). The maker of Ozempic and Wegovy is now paying two competing AI labs to work on the same underlying problem: how to get new drugs out of the lab and into patients faster. That detail matters more than either deal does by itself. It is a sign of how quickly AI drug discovery has stopped being a side experiment and become standard infrastructure across the pharmaceutical industry, and of how unsettled the question of which approach actually works still is, even among the AI companies themselves.

Novo Nordisk and Anthropic said the two companies will jointly identify drug-discovery problems raised by Novo's scientists and computational teams, then build targeted tools for what the companies called, in their own words, workflows to "support biological reasoning". As a first step, Novo will test Claude Science, Anthropic's dedicated research workbench, on specific research and development tasks.

Novo's chief executive, Mike Doustdar, framed the deal as part of a wider ambition "to become the world's most AI-driven healthcare company." He argued AI could "increase productivity in R&D and compress the path from research to marketed product," while also opening up "completely new scientific opportunities" for understanding human biology and how drugs work. Anthropic's co-founder and chief executive, Dario Amodei, went further, saying AI's growing capability "brings with it the potential to compress a century's worth of biological and medical breakthroughs into a decade". That is an aspiration from a chief executive with a direct commercial stake in the answer, not a measured forecast, and it is worth reading it that way. The companies also said the arrangement includes "robust data governance and human oversight" to keep AI use in line with Novo's compliance standards.

This is not Novo's first use of Claude. The company has already used it to automate the writing of clinical study reports, some running to 300 pages, cutting the process from around 12 weeks to about 10 minutes per report, according to Anthropic (Anthropic). Novo's Louise Lind Skov, director of content digitalisation, said the company had "consistently been one of the first movers when it comes to document and content automation in pharma development". The new deal moves Claude further upstream: from writing up results to being used in the scientific reasoning that comes before any results exist at all.

What is an AI workbench, and why does a drug company need one?#

Claude is a large language model, or LLM: a system trained on huge amounts of text and code that predicts, generates, and reasons about language, including scientific language such as chemical names, gene identifiers, and code for analysing lab data. Claude Science, which Anthropic launched in June 2026, is not a new or more capable model. Anthropic has said explicitly that it "runs the same Claude models already available to everyone today," with no special access (TechCrunch). What is new is the environment built around the model.

Ordinarily, a computational biologist doing this kind of in silico work, research carried out by computer simulation rather than at a lab bench, spends real time simply moving between separate databases and tools: one for gene sequences, another for protein structures, another for chemical compounds. Claude Science tries to put all of that in one place. A main AI assistant acts as something like a project manager, connecting to more than 60 scientific databases and tool libraries covering genomics, protein structure, and chemistry. It can then create "sub-agents", smaller AI assistants handling specific pieces of a task, the way a lead researcher might delegate parts of a project to different specialists.

A separate part of the system checks citations and calculations before anything is finalised. Anthropic also says each output, a protein structure diagram, say, comes bundled with the code and computing environment that produced it, plus a plain-language explanation, so other researchers can attempt to reproduce it (TechCrunch). That focus on reproducibility answers a real, long-standing complaint in computational biology, where results generated with one-off scripts and undocumented settings are notoriously hard for other labs to check.

Why this matters#

Novo's move fits a pattern rather than standing apart from one. NVIDIA and Eli Lilly opened a co-innovation lab in January 2026 backed by more than $1 billion over five years, aimed at closed-loop, AI-driven drug discovery (NVIDIA). Sanofi and Genmab have both adopted Claude for their own research and development work, according to Anthropic's published customer statements (Anthropic). Pharmaceutical companies have largely stopped treating generative AI as a side experiment run by one innovation team; it is becoming something closer to standard laboratory infrastructure, sitting alongside sequencing machines and lab information systems.

That shift reaches beyond the large companies making headlines. Anthropic is separately funding up to 50 academic research projects with Claude Science credits, focused on biomedical research and running from September to December 2026 (TechCrunch). If tools like this genuinely lower the cost of connecting genomic, structural, and chemical data, the biggest beneficiaries may not be large pharmaceutical companies with existing resources, but smaller academic labs and biotechs that previously could not afford dedicated bioinformatics teams.

For a company like Novo Nordisk, whose business rests on chronic disease drugs such as semaglutide, faster early-stage research could mean more candidate molecules reaching human trials sooner. Whether that translates into more approved medicines, rather than simply more candidates that fail a little later in the process, is a separate question, and one worth taking seriously rather than assuming away.

Critical analysis#

The strongest part of this announcement is also the easiest part to miss: Novo already has evidence, not just promises, that Claude saved real time on one specific, bounded task, clinical report writing. That is a much lower bar than "drug discovery," and pharmaceutical and AI companies both have a habit of blurring the two.

The evidence for the harder claim, that agentic AI tools meaningfully speed up the discovery and validation of new drugs, is thinner. A 2026 peer-reviewed review in the journal Pharmaceuticals found that global investment in AI for the life sciences has passed $100 billion since 2022, yet how much of that has actually improved clinical outcomes remains unclear. The review put the clinical attrition rate, the share of drug candidates that fail somewhere between entering clinical trials and winning regulatory approval, at roughly 90%, and found that gains in early-stage discovery speed have not consistently carried through to better late-stage success rates (Kumar, 2026, Pharmaceuticals). AlphaFold-style tools are very good at predicting static protein structure, the review notes, but weaker at modelling how proteins move, change shape, and interact with drug molecules in practice, which is often what actually decides whether a candidate drug works in the body.

A separate, cautionary note comes from an adjacent corner of clinical AI. In a September 2026 Nature Medicine comment, the Google research team behind AMIE, one of the best-known conversational diagnostic AI systems, argued that "trust in clinical AI cannot be benchmarked into existence" and can only be earned through rigorous prospective studies in real-world clinical settings, where the hardest lessons often concern the humans and systems around the AI rather than the model itself (Schaekermann et al., 2026, Nature Medicine). No equivalent prospective study of Claude Science in drug discovery exists yet. The Novo-Anthropic collaboration is, for now, a stated intention rather than a validated result.

One structural limitation is worth naming plainly. Claude Science's fact-checking layer is still built on the same underlying model it is meant to be checking, not an independent source of ground truth (TechCrunch). That does not make it useless: catching an inconsistency before publication is still genuinely useful. But it is a narrower safeguard than it might first sound.

Expert perspective: three different bets on the same problem#

Claude Science is not the only attempt to build AI tooling purpose-built for science, and the differences between the main approaches say more than any single product launch does. OpenAI took a narrower route in April 2026 with GPT-Rosalind, a model specifically fine-tuned for biological reasoning, but restricted to a research preview for vetted enterprise partners such as Amgen, Moderna, Thermo Fisher, and Novo Nordisk itself (TechCrunch). Google DeepMind is playing a third, different game again: rather than building a workbench on top of a general-purpose chatbot, it owns the underlying scientific models outright, such as AlphaFold for protein structure, and bundles them into its Gemini for Science platform alongside dozens of databases (TechCrunch).

So Anthropic is competing on breadth of access, available to anyone on a paid Claude subscription. OpenAI is competing on selective depth, a specialised model gated to a shorter list of partners. Google is competing on ownership of the underlying science itself, foundation models nobody else can call directly. That is a genuine three-way split in strategy, not just three companies chasing the same headline.

It also marks a change in ambition compared with the single milestone that made this whole field credible in the first place. AlphaFold, first released in 2020 and substantially expanded since, solved one well-defined problem: predicting a protein's three-dimensional shape from its sequence. Claude Science and its rivals are aiming much wider, at being a general working environment for many kinds of biological reasoning at once. Whether that broader ambition holds up, or whether it spreads the same underlying models too thinly across problems they were never specifically built to solve, is precisely what remains unresolved.

Key takeaways#

• Novo Nordisk is now paying two competing AI labs, Anthropic and OpenAI, to work on overlapping drug-discovery problems at the same time, rather than betting on a single AI vendor.
• Claude Science is not a new or more powerful AI model. It is a workbench connecting existing Claude models to more than 60 scientific databases and tools, plus a fact-checking layer built from that same underlying model.
• Novo's clearest evidence of impact so far is narrow: automating clinical study report writing, cutting the process from about 12 weeks to about 10 minutes. Evidence that AI meaningfully speeds up drug discovery itself is thinner.
• A 2026 peer-reviewed review found a clinical attrition rate of roughly 90% for drug candidates, and found no consistent sign that faster AI-assisted early discovery has improved that rate.
• Anthropic, OpenAI, and Google DeepMind are now running three distinct strategies in scientific AI: broad subscription access, narrow enterprise-gated access, and proprietary foundation models, and it is not yet clear which will win out.

Frequently asked questions#

What did Novo Nordisk and Anthropic actually announce? On 16 September 2026, the two companies said they would work together to address specific drug-discovery problems raised by Novo's scientists, starting with tests of Anthropic's Claude Science workbench on particular research and development workflows (BioSpace).

Is Claude Science a new, more powerful AI model built specifically for biology? No. Anthropic has said it runs on the same general Claude models available to any subscriber, with no special access for scientific use. What is new is the surrounding workbench: database connections, sub-agents, and a fact-checking step (TechCrunch).

Has Novo Nordisk dropped its partnership with OpenAI? No. Novo signed a separate, broad AI partnership with OpenAI in April 2026 covering drug discovery, manufacturing, and supply chains (CNBC). The Anthropic deal runs alongside it, not instead of it.

Does this mean AI is now discovering drugs on its own? No. The current use is more modest: helping scientists move faster through early-stage tasks such as literature review, hypothesis generation, and computational analysis. Human researchers still design experiments and interpret results, and Novo has said the arrangement includes human oversight requirements (BioSpace).

Is there published evidence that this kind of AI tool actually improves drug discovery outcomes? Not yet for Claude Science specifically. Independent evidence on AI in drug discovery more broadly is mixed: early-stage speed gains have not been shown to consistently improve the roughly 90% clinical attrition rate for candidates entering clinical development (Kumar, 2026, Pharmaceuticals).

How is this different from what Google DeepMind and OpenAI are doing? OpenAI's equivalent tool, GPT-Rosalind, is a specialised model restricted to vetted enterprise partners. Google DeepMind owns its underlying scientific models, such as AlphaFold, and bundles them into its own Gemini for Science platform. Claude Science instead offers broader access built on general-purpose models (TechCrunch).

Can researchers outside large pharmaceutical companies use Claude Science? Yes. It is available in beta to anyone on a Claude Pro, Max, Team, or Enterprise subscription, and Anthropic is separately funding up to 50 academic biomedical research projects with usage credits (TechCrunch).

When might this kind of partnership actually change which drugs reach patients? Given that drug development typically runs a decade or more from early research to approval, and given the current lack of published results specific to this collaboration, any measurable effect on approved medicines is more likely years away than months.

Glossary#

Large language model (LLM): an AI system trained on very large amounts of text and code to predict and generate language, including specialised scientific and technical language.

Agentic AI / AI agent: an AI system set up to carry out multi-step tasks with some autonomy, such as choosing which tool or database to query next, rather than just answering a single question.

Foundation model: a large, general-purpose AI model, such as Claude or GPT, that can be adapted or built upon for many different specific uses, including scientific ones.

AI workbench: software that provides a single working environment connecting an AI assistant to multiple external tools, databases, and file formats relevant to a particular field, such as biology or chemistry.

In silico: research carried out by computer simulation or modelling, from the Latin for "in silicon", as distinct from research in a lab (in vitro) or in a living organism (in vivo).

Clinical attrition: the rate at which drug candidates fail during clinical trials or regulatory review rather than reaching approval and market use.

Reproducibility (computational): the ability of another researcher to obtain the same result by rerunning the same code, data, and computing environment used in the original analysis.

Biological reasoning: forming and testing hypotheses about how genes, proteins, cells, or organisms behave, as distinct from simply retrieving known facts.

References#

1. Novo Nordisk and Anthropic. "Novo and Anthropic will collaborate to advance drug discovery with Claude." BioSpace / GlobeNewswire, 16 September 2026. https://www.biospace.com/press-releases/novo-and-anthropic-will-collaborate-to-advance-drug-discovery-with-claude
2. Anthropic. "Advancing Claude in healthcare and the life sciences." Anthropic News, 11 January 2026. https://www.anthropic.com/news/healthcare-life-sciences
3. Bellan, R. "Anthropic's Claude Science bets on workflow, not a new model, to win over scientists." TechCrunch, 30 June 2026. https://techcrunch.com/2026/06/30/anthropics-claude-science-bets-on-workflow-not-a-new-model-to-win-over-scientists/
4. CNBC. "Novo Nordisk partners with OpenAI as AI drug discovery hopes mount." CNBC, April 2026. https://www.cnbc.com/2026/04/14/novo-nordisk-openai-ai-drug-discovery-healthcare-nvo.html
5. Kumar, S. "AI in Drug Discovery: Clinical Failures, Regulatory Reality, and the Validation Crisis Behind the Hype." Pharmaceuticals, 19(6), 916 (2026). Peer-reviewed. https://www.mdpi.com/1424-8247/19/6/916
6. Schaekermann, M., Palepu, A., Rodman, A. et al. "Prospective evidence for conversational medical AI is hard, but non-negotiable." Nature Medicine, comment, published 14 September 2026. https://www.nature.com/articles/s41591-026-04639-5
7. NVIDIA Newsroom. "NVIDIA and Lilly Announce Co-Innovation AI Lab to Reinvent Drug Discovery in the Age of AI." January 2026. https://nvidianews.nvidia.com/news/nvidia-and-lilly-announce-co-innovation-lab-to-reinvent-drug-discovery-in-the-age-of-ai