OpenAI has published a case study claiming that GPT-5 Pro resolved a three-year-old question about how glucose shapes T cell specialization, an immunology puzzle that had stalled in the lab of Derya Unutmaz at the Jackson Laboratory for Genomic Medicine. The framing is unmistakable: the company isn’t selling a chatbot, it’s selling a scientific collaborator.
The set-piece moment, recounted by Unutmaz in late 2025, was a test. He asked GPT-5 Pro to simulate the outcome of a wet-lab experiment he had already run but not yet published, involving anti-CD19 CAR memory CD8+ T cells engineered to target lymphoma. The model predicted the result correctly.
“That was the moment that I felt like, okay, these models have now come to a point where they really, truly understand,” Unutmaz said.
A companion technical report posted to arXiv (“Early science acceleration experiments with GPT-5”) fills in the mechanism. GPT-5 Pro distinguished a glycolysis blockade from impaired N-linked glycosylation, proposed the IL-2 receptor pathway as the driver, and predicted that memory rather than naïve T cells would carry the effect. It then produced what the authors call a clean decision tree of follow-up experiments, including a mannose rescue to restore glycosylation, plus metabolic and epigenetic assays. The model also correctly predicted that brief exposure to 2-deoxyglucose during priming would lower exhaustion markers PD-1 and LAG-3 while preserving cytotoxic potential.
The arXiv authors argue the chief contribution wasn’t a single insight but the elimination of dead ends, separating selection effects from programming effects and sparing the lab from “attractive but potentially unnecessary experiments, which would have wasted many months of testing.” A second OpenAI case study describes Unutmaz using GPT-5.5 Pro to chew through a 62-sample, 28,000-gene expression dataset in work he said would’ve taken his team months.
The case study lands inside a larger campaign. GPT-5, OpenAI’s first “unified” model, shipped in August 2025. By January 2026, MIT Technology Review reported that OpenAI had stood up an OpenAI for Science unit explicitly to court academic users, putting it into direct competition with Google DeepMind, whose AlphaFold remains the canonical example of AI doing legible scientific work.
“I think 2026 will be for science what 2025 was for software engineering,” OpenAI executive Kevin Weil told the magazine. The line reads less as prediction than as positioning. AlphaFold gave DeepMind the lore; OpenAI now wants its own.