The AI 'Doing Science' Is Mostly a Database Lookup. Zero AI-Discovered Drugs Have Been Approved. The Gap Between Those Two Facts Is the Story.
On VirBench, Claude scored ~17% until researchers bolted on a deterministic tool that looks answers up in NCBI — then it hit 99.7%. On NatureBench, the best of ten frontier agents beat published state-of-the-art on just 17.8% of tasks, winning by 'methodological translation,' not invention. Even OpenAI's flagship science model, GPT-Rosalind, calls AlphaFold as an external tool. The verifiable gains in AI-for-science come from the deterministic scaffolding, not the model's reasoning — and the durable moat is whoever owns the data pipeline and regulatory path, not whoever ships the smartest model.
Here are two of the cleanest numbers in AI right now, and they're both about science. On VirBench, a virology benchmark, Claude scored around 17% — until researchers bolted on a deterministic retrieval tool that simply looks answers up in NCBI's databases, at which point accuracy jumped to 99.7%. The paper's own title says deterministic access "enables robust agentic scientific discovery." Read it the other way and it's more honest: the lookup is the discovery. On NatureBench, the best of ten frontier agents beat published state-of-the-art on just 17.8% of Nature-family tasks, and the authors found it won mostly through "methodological translation" — turning scientific problems into ordinary supervised-prediction problems — "rather than through genuine scientific invention." Those are the numbers sitting underneath every headline announcing that AI now does science.
It mostly doesn't. More precisely: the verifiable gains in AI-for-science are coming from the deterministic tools wrapped around the models, not from the models' scientific reasoning — and the whole industry is structured to keep that distinction blurry. Once you start looking for the tell, it's in every launch. The cleanest example is OpenAI's own flagship science model.
GPT-Rosalind, launched in April, scores 0.751 on the BixBench bioinformatics benchmark — first place, comfortably ahead of the general-purpose models. And to do its job it calls AlphaFold 3 as an external tool, alongside 50-plus public databases. Sit with that architecture: even OpenAI, building a model specifically for biology, didn't put the science in the weights. It built an orchestrator that drives specialized deterministic tools, because the tools are where the reliability lives. The BixBench crown, meanwhile, is doing rhetorical work it hasn't earned — BixBench measures bioinformatics, and there is no validated mapping from a bioinformatics leaderboard to the things drug discovery actually needs: target identification, lead optimization, ADMET prediction. A benchmark win is not a clinical outcome, and the launch coverage treats them as the same currency. Amgen, Moderna, and Novo Nordisk signed on as partners, which is real money interested, but the financial structure of those deals is undisclosed, so what they're actually buying is anyone's guess.
Two questions cut through the entire field. First: is there a single validated, peer-reviewed, independently replicated result attributable to a model's reasoning — not its tooling or its training data? As best I can tell, no. The strongest candidates all fail the test. The Chan-Lam chemistry improvement was real and experimentally validated, but it happened inside Molecule.one's automated wet lab, which means the credit splits heavily toward the infrastructure, and it was published by the vendor, not independently refereed. The celebrated GPT-5 Pro immunology "prediction" is N=1, self-published, with the obvious contamination question — did it reason, or did it generalize from the literature it was trained on? — left unaddressed. The AI-designed drug that reached Nature Medicine was a molecular-design platform result, not language-model reasoning, and it's only at Phase IIa. Second question: does any benchmark in this space have a published mapping to a clinical or experimental milestone? Also no. BixBench, NatureBench, VirBench all measure proxies, and a proxy is exactly what gets quietly upgraded to "breakthrough" in a press release.
Which brings up the fact that should anchor every one of these stories and almost never does: no AI-discovered drug has cleared Phase III or won FDA approval. Roughly fifteen are entering Phase III out of more than two hundred in clinical development, and the count of approvals is zero. The first one is projected for 2026 or 2027 with something like 60% analyst confidence, which is to say it's plausible and unproven. And notably, Big Pharma is conspicuously absent as the originator of these launches — the molecules in the pipeline mostly come from AI-native biotechs, not from the incumbents whose names appear on the partnership slides.
Here's the part that's actually a business story rather than a debunk. If the model isn't the moat, what is? Ownership of the deterministic tools, the governed data, and the regulatory path. Watch where the sophisticated players are planting flags. Microsoft and Mayo Clinic built a frontier healthcare model that Mayo owns — trained on Mayo's longitudinal clinical data, distributed through Azure, with the hospital controlling licensing and liability while Microsoft owns the pipes. Databricks and NVIDIA shipped Genesis Workbench, a drug-discovery environment with no external API at all, so a pharma company's proprietary IP never leaves its own governed catalog — purpose-built to defeat the security objection that blocks per-token model APIs. Anthropic's "AI for Science" event was a positioning move, not a model launch: named pharma references, a marquee scientist's first public appearance, a land-grab for the pipeline and the relationships ahead of any proven cure. The capability race is the show. The data-and-regulatory-pipeline race is the business.
The honest caveats matter, because the cynical read can curdle into "AI is useless for science," which is wrong. Deterministic-tool-plus-model is genuinely valuable — taking a virology task from 17% to 99.7% is a real capability, even if the credit belongs to the scaffolding. The Chan-Lam result happened. And an AI-discovered drug really might clear Phase III in the next year or two, which would be the first piece of evidence that closes the benchmark-to-bedside gap rather than papering over it. So this isn't a prediction that the field fails. It's an argument about where the value is and what counts as proof. My read: the durable margin accrues to whoever owns the pipeline, not whoever tops the next leaderboard, and the "smart model finds cures" narrative is going to keep outrunning its evidence until a regulator, not a benchmark, signs off. What would change my mind is specific and binary — a validated, independently replicated result attributable to model reasoning alone, plus an AI-discovered drug through Phase III. Until both of those exist, the thing to watch isn't the next BixBench score. It's the first Phase III readout and the FDA framework that will or won't accept it. That's the milestone. Everything before it is scaffolding wearing a lab coat.