There are 200+ AI-discovered drugs in trials and zero FDA approvals. 2026 is when that finally gets tested.

Here is the state of AI drug discovery in mid-2026, stated as plainly as the evidence allows: there are more than 200 AI-discovered drugs in clinical development, roughly 15 of them in Phase 3, and zero have received FDA approval. All three of those numbers matter, and the third one matters most. This is a field that has absorbed extraordinary capital, generated extraordinary activity, and — as of today — produced no finished medicine. Any honest accounting starts there, because everything else in this story is provisional until that number moves off zero.

That is not a dismissal — it is the honest denominator, and denominators are what this beat exists to enforce. And 2026 is the year the denominator finally starts to resolve, because industry estimates suggest 15 to 20 AI-associated programs may enter pivotal Phase III trials this year. To understand why that matters, you need to understand what Phase III actually is: the large, expensive, multi-year test of whether a drug improves human outcomes in a population big enough to mean something. It is where the field's optimism has always deferred its reckoning, because Phase III is where drugs go to fail — roughly half of all conventional drugs that reach it don't survive it, and that's after decades of refinement in how candidates get chosen. Mouse studies cannot substitute for it. Biomarkers cannot substitute for it. Elegant molecular design cannot substitute for it. Which means we are about to learn, with real human data, whether AI-driven discovery produces medicines that work in people — or merely candidates that looked wonderful on the way in.

What actually counts as progress

The milestones to date are real, and precision about them is a form of respect — for the science and for readers trying to calibrate. AI drug discovery cleared its first Phase IIa test in 2025, the moment one industry gathering described as the shift 'from speculation to evidence.' Relay Therapeutics' zovegalisib has received FDA Breakthrough Therapy Designation and is now in Phase 3 for metastatic breast cancer — arguably the field's most advanced flag-bearer. Companies like GT Biopharma are threading AI across the engineering of tumor-targeting therapies, with candidates aimed at pre-IND development and early trial starts. Each of these is a genuine, creditable step. None of them is an approval. And the distance between 'Breakthrough Therapy Designation' and 'approved medicine' is still measured in years, attrition, and the specific humility that oncology enforces on everyone who works in it.

Phase III is where drugs go to fail. That's exactly why it's the only test that will tell us whether AI discovery is real.

The bull case and the bear case — both still live

The bull case runs like this: drug discovery's front half — finding and optimizing candidate molecules — has historically taken four to six years, and AI demonstrably compresses it. The 200+ candidates now in the clinic are the leading edge of a wave that simply hasn't had time to reach the finish line, because a drug discovered in 2021 is only now arriving at its pivotal trials. On this reading, zero approvals isn't a warning sign; it's an artifact of the calendar, and the wave breaks over the next three years.

The bear case is quieter and more uncomfortable: discovery was never actually the bottleneck. The brutal economics of pharma live in clinical failure — candidates that looked perfect until they met human biology at scale. There's an old story every medicinal chemist knows: the industry has been burned before by technologies that promised to fix discovery, from combinatorial chemistry in the 1990s to high-throughput screening after it. Each one produced more candidates, faster and cheaper — and the Phase III failure rate barely moved, because generating candidates was never the hard part. If AI's contribution is more shots on goal against the same goalkeeper, the revolution will have been a cost improvement rather than a success-rate improvement. That would still matter — cheaper discovery is not nothing — but it would be a far smaller something than the valuations assume.

Both cases are fully live, and anyone who tells you otherwise is selling a position. What makes 2026 different is that, for the first time, the question has an expiration date: this year's and next year's Phase III readouts are the answer arriving.

What to watch, and how to watch it honestly

The first FDA approval of an AI-discovered drug, whenever it comes, will be a genuine landmark — watch for it, but resist pre-writing it, because a single approval will be seized on as vindication for an entire methodology, which one data point cannot provide. The more informative signal is the cohort: as the 15–20 pivotal programs read out, does the AI-discovered class clear Phase III at a rate meaningfully better than the historical baseline? That comparison — rate versus rate, not molecule versus headline — is the entire question. And watch the failures as honestly as the wins: an AI-discovered drug that fails Phase III is data about the method, not just the molecule, and a field that only publicizes its survivors is a field asking not to be measured. We'll grade this story against real outcomes as they land, in public, the same way we grade everything else.