Three years. That’s how long Dr. Derya Unutmaz, a prominent immunologist at the Jackson Laboratory, had been staring at a biological puzzle he couldn’t crack. Then he spent a few sessions with GPT-5 Pro and got the answer. If that doesn’t make you stop and think about what AI is actually becoming in the lab, I’m not sure what will.
OpenAI published the full story on June 23, 2026, and it’s worth reading beyond the headline. This isn’t a “AI assists researcher” fluff piece. This is a scientist describing a genuine intellectual breakthrough — one that had stumped him and his team for years — attributed directly to a conversation with a language model. The implications for how we do science are hard to overstate.
What Was the Mystery, Exactly?
Unutmaz’s research centers on T cells — the immune system’s frontline soldiers. For years, his lab had been tracking a specific and puzzling behavior: certain T cell populations were acting in ways that didn’t fit established immunological models. The data was real, reproducible, but the explanation? Nowhere to be found in the literature.
This is the part that gets lost in headlines. Unutmaz wasn’t asking GPT-5 to run a Google search or summarize a Wikipedia article. He was presenting it with proprietary experimental data, years of contextual research, and asking it to reason through biological mechanisms that haven’t been cleanly documented anywhere. That’s a fundamentally different ask than “explain what a T cell is.”
What GPT-5 Pro apparently did — and this is where it gets genuinely interesting — was synthesize patterns across immunology, cell biology, and adjacent fields in a way that pointed toward a mechanistic explanation Unutmaz and his team hadn’t considered. The model wasn’t just retrieving. It was connecting dots across disciplines in a way that took a human expert three years to not do.
Unutmaz described the experience as less like using a search engine and more like collaborating with “a brilliant colleague who has read everything.” That framing matters. It’s not a tool. It’s starting to function like a research partner.
Why T Cells, and Why Does This Matter Beyond One Lab?
T cells sit at the center of some of the most important medical challenges we’re dealing with right now. Understanding their behavior — particularly when they malfunction or behave unexpectedly — is foundational to progress in three areas:
- Cancer immunotherapy: CAR-T therapies and checkpoint inhibitors depend entirely on manipulating T cell behavior. Any new insight into how T cells make decisions, activate, or exhaust changes what’s therapeutically possible.
- Autoimmune disease: Conditions like lupus, multiple sclerosis, and rheumatoid arthritis are essentially T cells attacking the body’s own tissue. A clearer mechanistic picture helps researchers design better interventions.
- Aging and immune decline: Unutmaz’s lab has done significant work on immunosenescence — how the immune system ages. T cell behavior is central to that research too.
So when Unutmaz says this breakthrough could “support cancer and autoimmune research,” he’s not being vague. He’s pointing at a very specific set of downstream applications that could eventually affect millions of patients. The question is how far away those applications are — and whether AI can keep accelerating the path from insight to intervention.
GPT-5 Pro as a Scientific Instrument
Here’s the thing people keep getting wrong about GPT-5 in research contexts: they assume it’s being used as a fancy literature review tool. Read the papers, summarize the findings, save the researcher some time. That’s real and useful, but it’s not what happened here.
What Unutmaz describes is something closer to hypothesis generation through analogical reasoning. GPT-5 Pro — the most capable tier of OpenAI’s current flagship model — has enough breadth across scientific domains that it can surface connections between, say, metabolic signaling in one cell type and receptor behavior in another, even when those connections haven’t been explicitly drawn in published literature.
This is the capability that separates GPT-5 from earlier models in a research context. GPT-4 was genuinely useful for writing assistance, code generation, and literature synthesis. It struggled with deep, multi-step scientific reasoning that required holding many variables in tension simultaneously. GPT-5 Pro, with its extended context window and improved reasoning architecture, handles that kind of sustained intellectual work much better.
Compare that to what Google’s Gemini 2.5 Ultra or Anthropic’s Claude Opus 4 can do in similar research contexts — both are capable, both are being used in labs. But the Unutmaz case is a pointed example of OpenAI making a credible claim that GPT-5 Pro is doing something qualitatively different. Not just faster or cheaper — actually different in kind.
We’ve seen similar dynamics play out in other domains. Our coverage of the GPT-5-powered AI chemist improving drug synthesis showed the model operating in a similarly generative role — not just assisting chemists but actively proposing novel synthesis pathways. The immunology case fits the same pattern. And OpenAI’s LifeSciBench benchmarks have been trying to quantify exactly this kind of capability — how well can AI models reason across complex life science problems?
The Reproducibility Question Nobody’s Asking Loudly Enough
I want to flag something that tends to get glossed over in these announcements: we’re hearing one scientist’s account of a breakthrough. The mechanistic explanation GPT-5 offered hasn’t — as of this writing — been independently validated, peer-reviewed, or published in a journal.
That’s not a knock on Unutmaz or on OpenAI. It’s just where things are. Scientific breakthroughs, even genuine ones, take time to move through validation. The T cell behavior explanation could turn out to be exactly right. It could also turn out to be a plausible-sounding hypothesis that doesn’t survive experimental scrutiny.
This is actually the critical question for AI in research right now: how good are these models at being confidently wrong in ways that sound convincing? GPT-5 is significantly better than its predecessors at signaling uncertainty, but it’s not perfect. A researcher with deep domain expertise — like Unutmaz — is well-positioned to evaluate a hypothesis critically. A less experienced researcher might accept a sophisticated-sounding explanation that’s subtly off.
That’s not a reason to stop using AI in research. It’s a reason to use it carefully, with human expert review baked into the process. The Unutmaz case seems to have that — he’s not replacing his lab with GPT-5, he’s using it as one input into a longer scientific process.
What This Means for Researchers and Institutions
If you’re working in biomedical research — or funding it — the practical takeaways from the Unutmaz case are actually pretty concrete:
- Domain expert + powerful model = real leverage. This only worked because Unutmaz knew enough to ask the right questions, frame the problem correctly, and evaluate the answer critically. AI isn’t replacing immunologists. It’s making excellent immunologists faster.
- Access matters. GPT-5 Pro isn’t cheap. The most capable tier of OpenAI’s offering sits behind a meaningful price point. Research institutions that can afford broad access will move faster than those that can’t. That’s a funding and equity question the field hasn’t fully reckoned with.
- Interdisciplinary reasoning is where AI earns its keep. The more siloed a research question is, the less useful a general model is. But when a problem requires synthesizing across disciplines — which many hard biological questions do — that’s where models like GPT-5 Pro have a genuine edge over any single human expert.
- Documentation is underrated. Researchers using AI tools should be logging their interactions systematically. The conversation that solved Unutmaz’s mystery is scientifically interesting in its own right — what was asked, how the model framed its response, what follow-up questions moved the needle.
We’re also watching this intersect with a broader OpenAI push into scientific and institutional use cases. The OpenAI Partner Network’s $150M investment program is specifically designed to accelerate exactly this kind of applied research deployment — getting GPT-5 into the hands of domain experts in structured, supported ways.
FAQ
What exactly did GPT-5 Pro do to help solve the immunology mystery?
GPT-5 Pro helped Dr. Unutmaz by reasoning across multiple scientific domains simultaneously to identify a mechanistic explanation for anomalous T cell behavior his lab had observed for three years. Rather than simple literature retrieval, the model synthesized cross-disciplinary connections that hadn’t been explicitly documented, functioning more like a collaborative research partner than a search tool.
Does this mean AI can now do independent scientific research?
Not quite. The Unutmaz case involved a highly experienced immunologist who framed the problem, provided rich context, and critically evaluated the model’s output. GPT-5 Pro generated a compelling hypothesis — but validating it still requires traditional experimental science. AI is accelerating the hypothesis generation phase, not replacing the experimental validation that follows.
How does GPT-5 Pro compare to other AI models for research use?
GPT-5 Pro, Claude Opus 4, and Gemini 2.5 Ultra are all being used in research settings. GPT-5 Pro’s extended reasoning capabilities and large context window give it an edge in tasks requiring sustained multi-step analysis across large bodies of information. That said, different models have different strengths, and many research teams are using multiple models depending on the task.
When will this research be published?
As of June 2026, the findings described by Unutmaz have not been published in a peer-reviewed journal. OpenAI’s announcement describes the experience and the insight, but the scientific validation process — which includes peer review and likely additional experiments — is presumably ongoing. Watch for publications from the Jackson Laboratory’s Unutmaz lab in the coming months.
The Unutmaz case will almost certainly not be the last story like this. Labs around the world are running similar experiments with GPT-5 right now, and a handful of them are probably sitting on results they haven’t announced yet. The more interesting question isn’t whether AI can help solve scientific mysteries — it’s how we build the institutional infrastructure to make that collaboration systematic, reproducible, and equitably accessible across the research community. That’s the harder problem, and nobody’s solved it yet.
