When AI Companies Become Pharma Startups

Anthropic announced something unusual on July 4, 2026: the company is launching its own drug discovery programs aimed at neglected diseases. Not diseases that are medically unsolvable, but diseases that traditional pharmaceutical companies consider unprofitable to pursue[1].

The announcement came during an event for Claude Science, Anthropic's new AI research workspace. The company says the move aligns with its nonprofit mission and will help it build better AI models through firsthand experience with the drug discovery pipeline[2].

This is a strange and interesting pivot. Anthropic is primarily known as an AI safety company. Now it is also a preclinical drug development lab. The framing is that building drug discovery tools requires dogfooding, that you cannot build good software for an industry without working inside it. There is logic to that. But it also means an AI company with no wet lab, no clinical trial infrastructure, and no regulatory experience is now in the pharmaceutical business.

What They Actually Showed

Anthropic demoed a few early results at the event. A researcher at UCSF used Claude Science to spot a viral contamination in minutes that his team had missed for an entire year. Claude analyzed 100 rare genetic diseases in under an hour and flagged 32 candidates for computational screening[2].

These are computational results, not clinical ones. They are impressive as speed metrics but they are early stage. The gap between "flagging a candidate" and "having a drug" is enormous. Anthropic is positioning itself at the very beginning of that pipeline, in preclinical research, which is the part where AI can plausibly add the most value.

The Novartis Perspective

Novartis CEO Vas Narasimhan appeared at the same event and offered a more grounded view. He broke drug development delays into three categories: information latency, operational latency, and biological latency[3].

AI tools could meaningfully cut the first two, which account for roughly 40 percent of total development time. Biological latency, the time needed for animal testing, cell models, and human clinical trials, will not shrink much regardless of how good the software gets. Narasimhan's estimate: development timelines could come down from twelve years to seven or eight. Success rates could potentially double from 8 percent to 16 percent.

These are modest gains by tech industry standards but massive by pharma standards. The big pharmaceutical companies collectively spend $150 to $200 billion per year on R&D and have produced only 800 to 1'000 drugs in 120 years[3]. Even a small improvement in the hit rate translates to enormous value.

The Competitive Landscape

Anthropic is not alone here. Alphabet's Isomorphic Labs, built on DeepMind's AlphaFold protein structure prediction, is already preparing for its first human trials with AI-designed drugs. John Jumper, the Nobel laureate who co-developed AlphaFold, recently left Google for Anthropic[4]. That hiring move makes more sense in retrospect, Anthropic was clearly building toward a science vertical.

OpenAI has been moving into healthcare too, launching ChatGPT Health earlier in 2026 with medical record integration. Google DeepMind introduced an AI Co-Clinician built around triadic care, where AI agents support patients while physicians retain clinical authority[5].

The pattern is clear: every major AI lab is looking at biology and medicine as the next frontier. The reasons are obvious. Drug discovery is a combinatorial optimization problem at scale, which is exactly the kind of task where large language models and their descendants could plausibly outperform humans. And the market is enormous, underserved, and scientifically legitimate.

The Skeptical View

Independent experts urge caution. Catherine Pope of the University of Oxford described existing AI clinical results as "a piece removed from the messy, complex, human world of everyday healthcare"[6]. The gap between computational analysis and actual patient outcomes is where most medical AI projects stall. Spotting a contamination in a dataset is not the same as developing a safe, effective drug that passes through regulatory approval.

There is also a structural question about whether an AI company should be running drug programs at all. Anthropic's argument is that neglected diseases are precisely the ones where market incentives fail and where a nonprofit-mission-driven actor could add genuine value. That is a reasonable argument. It is also the kind of argument that sounds great in a press event and gets tested only over a decade of preclinical and clinical work.

Why This Matters

What makes this worth paying attention to is not the specific drug candidates Anthropic might produce. It is the signal it sends about where AI companies see themselves going. The first wave of AI was search and recommendation. The second wave was conversation and code generation. The third wave might be molecular design and drug discovery, and the companies building the models want to be in the pipeline directly, not just selling tools to pharma.

Whether that vertical integration works is an open question. But the land grab has started.