There’s a kind of AI story that sticks with you more than any benchmark table. OpenAI told one of those on June 23: how GPT-5 Pro helped immunologist Derya Unutmaz solve a puzzle that had been sitting in his drawer since 2022.
The mystery in the test tube
Back in 2022, Unutmaz — a professor at The Jackson Laboratory and the University of Connecticut — wanted to understand how glucose shapes the development of T cells. T cells are the immune cells that fight viruses, kill cancer cells, and tell healthy cells from threats. Depending on how they specialize, they influence cancer, autoimmune disease, and infection.
His team exposed young T cells either to a low-glucose environment or to a glucose-like molecule called deoxyglucose, which blocks a cell’s ability to use glucose. In theory, both conditions should have produced similar results — in both cases the cell is starved of energy. They didn’t. The deoxyglucose cells overwhelmingly became inflammatory cells; the low-glucose ones only partly. A lack of energy alone couldn’t explain that. The lab couldn’t crack it, so they shelved the experiment and moved on to more urgent work.
What GPT-5 saw
When GPT-5 Pro arrived in late 2025, Unutmaz pulled the old data back out and fed it to the model. The answer landed: deoxyglucose interferes with building a protein called IL-2 — and IL-2 is exactly what normally stops a T cell from turning into the inflammatory Th17 type. Remove the protein, remove the brake. “GPT-5 came up with this really remarkable insight that, retrospectively, makes perfect sense,” Unutmaz says. Nobody in the lab had seen the connection.
Then came the real test. Unutmaz asked the model to predict the outcome of a different, still-unpublished experiment — the ability of certain CD8+ cells to kill lymphoma cells. GPT-5 Pro predicted the effect correctly. It couldn’t have gotten the answer off the internet; the results weren’t published anywhere. “That was the moment I felt like, okay, these models have now come to a point where they really, truly understand.”
My take
What grabs me about this story isn’t the “AI cures cancer” drama — that would be overblown. It’s the sober part: the model didn’t invent data and didn’t replace a human. It produced a hypothesis that an expert still had to recognize, test, and weigh. Without Unutmaz’s knowledge, the insight would have been worthless — he says so himself.
That, to me, is the realistic version of AI in science: not an oracle, but a collaborator that changes the tempo. Weeks or months of lab work fall away because you know faster which experiment is worth running. The flip side belongs in the picture too: what speeds up hypotheses for cancer research also lowers barriers for misuse. Both are true. But cracking a three-year-old mystery on a Tuesday — that’s pretty impressive.
This article touches on medical research. It is not a substitute for professional or medical advice.
Sources: OpenAI: How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery