OpenAI just entered the AI drug discovery trade

OPENAI LAUNCHED GPT-Rosalind on Thursday as a research preview for life sciences, with Amgen, Moderna, and the Allen Institute as founding customers. The share prices that moved on the news belonged to none of them. Recursion Pharmaceuticals fell more than 5%, as did Schrödinger; Charles River Laboratories dropped 2.6%, IQVIA 3.2%. OpenAI shipped a research model for biology. The market read it as a category obituary.

The four companies that sold off share a common investor pitch: specialized AI built for biological data, wet-lab integration, and proprietary datasets accumulated over years. Recursion runs 2.2 million cellular experiments a week on an Nvidia-built supercomputer, and last July paid $688 million for Exscientia to assemble what its executives called the broadest end-to-end AI drug discovery platform on the market. Schrödinger pairs physics-based molecular simulation with machine learning; its lead molecule, zasocitinib, has now reached Phase III trials. Charles River and IQVIA sit a layer above, selling preclinical services and trial logistics to the companies doing the actual discovery. The pitch throughout has been that domain specialization compounds into a durable moat, and that generalist AI, however capable, cannot traverse the chasm between predicting a protein fold and identifying a clinically useful molecule.

Double helix, single exit

But specialization is a moat only if generalists cannot reach the frontier. GPT-Rosalind is a research preview built on OpenAI's newest internal models, shipped alongside a free Life Sciences plugin that gives Codex users access to more than 50 scientific tools and databases. It is not a finished drug-design system; OpenAI concedes the model cannot, on its own, devise new treatments, and framed it to reporters as a research partner rather than a researcher. That is, however, precisely the framing that makes the valuation reset sting. The market was not pricing the product in front of it. It was pricing the thesis behind the category: that "AI for drug discovery" as a commercial proposition survives a frontier lab turning up to the work, even casually, even early, even with a research preview.

The signal was already there for anyone willing to read it. Recursion traded above $41 in July 2021; today it sits at $3.60, a decline of more than 90%, and Nvidia liquidated its entire stake in February. Exscientia, once an industry darling, lost its Bayer partnership before being swallowed at a small fraction of the market capitalization it commanded at its post-IPO peak. BenevolentAI's lead program failed. Absci, Relay, and AbCellera have each given back most of their pandemic-era gains. Between them, the public AI-biotech pure-plays have delivered zero approved drugs and a run of partnerships that in hindsight look less like validation and more like data-licensing agreements at favorable prices for the incumbents writing the checks.

None of that is proof that AI cannot design drugs. It is proof that domain-specialized AI, operated as a standalone company on public markets, cannot capture the value of doing so. Insilico Medicine's rentosertib, the first end-to-end AI-designed drug to deliver positive Phase II results, belongs to a private company. So does Isomorphic Labs, which raised $600 million from Thrive Capital last year on top of nearly $3 billion in potential milestones from Eli Lilly and Novartis. Xaira Therapeutics launched in 2024 with $1 billion, the largest seed round in biotech history, behind David Baker's Nobel-laureate protein-design group. Generate:Biomedicines filed for a Nasdaq IPO in February, weeks after dosing the first Phase III patient for an AI-designed antibody. Each survivor is anchored in something a foundation model cannot synthesize from the open literature: an Alphabet parent with proprietary AlphaFold training, a Nobel laureate's protein lab, a generative-biology platform with a clinical molecule already in trials. None of them is a wrapper.

The pattern is not specific to biology. The vertical AI wrappers that rose on top of GPT-3 and GPT-4 (Jasper for copywriting, a flotilla of "AI for [vertical]" apps built on thin API stacks) have been getting squeezed since ChatGPT itself absorbed most of what they did. Biology was supposed to be different because biology has physics, wet labs, and regulation. What GPT-Rosalind establishes is that those barriers slow the frontier lab down by a research preview's worth of iteration, not by a category's worth. Anthropic, which recently acquired Coefficient Bio, a months-old generative-protein startup staffed with Genentech's former Prescient Design team, is building toward its own. Google already has Isomorphic in-house. The middle layer — vertical-AI wrappers on top of readable datasets — is filling in fast, and getting priced out faster.

Amgen and Moderna will use GPT-Rosalind; so will everyone else in pharma. Five years from now, the survivors with any premium left will be the ones that own a molecule, a wet lab, or a frontier model of their own. Thursday's verdict, delivered in real time as OpenAI was still briefing reporters, was that the list had grown shorter by four.