GPT-5 Pro Helped Crack a 3-Year Immunology Mystery About T Cells

Immunologist Derya Unutmaz used OpenAI's GPT-5 Pro to resolve a question about T cell behavior that had stalled his research for three years, according to a post on the OpenAI blog. The result, which OpenAI says could inform work on cancer and autoimmune conditions, is one of the clearest public examples so far of a large language model contributing meaningfully to ongoing scientific research rather than just summarising existing literature.

What happened

Derya Unutmaz, an immunologist, had been sitting on an unresolved question about T cell behavior for three years. T cells are a critical part of the immune system: they identify and attack pathogens and abnormal cells. Understanding exactly how they behave, and why they sometimes behave unexpectedly, is central to developing better treatments for cancer and autoimmune diseases.

According to OpenAI, Unutmaz worked with GPT-5 Pro and came away with new insight into that T cell puzzle. OpenAI published the story on its blog as an illustration of GPT-5’s capability in high-stakes scientific domains.

Why it matters

This is not a story about AI writing a paper or generating a hypothesis list. A working scientist with a specific, years-old problem says he got a meaningful answer by working with GPT-5 Pro. That is a different claim than most AI-in-research stories, which tend to be about speed or productivity gains on routine tasks.

T cell research sits at the heart of some of the most active areas in medicine right now:

• Cancer immunotherapy: treatments that direct T cells to attack tumours depend on precise understanding of T cell activation and exhaustion.
• Autoimmune disease: conditions like lupus, multiple sclerosis, and rheumatoid arthritis involve T cells attacking the body’s own tissue.

If GPT-5 Pro can genuinely accelerate the kind of reasoning that moves a stalled research question forward, the implications for drug discovery timelines and research costs are real. That said, OpenAI is the source here, so the framing is promotional. Independent replication and peer review of whatever Unutmaz found will matter a great deal.

Our take

We should be honest about what we know and what we do not. OpenAI published this story. We do not have the underlying research, the specific question Unutmaz asked, or the output GPT-5 Pro gave him. What we have is a case study from the company selling the product.

That does not mean it is false. It means we should hold it at arm’s length until there is more detail. The more interesting question for most people reading this is not whether one immunologist had a good session with GPT-5 Pro. It is whether that kind of deep, domain-specific reasoning is becoming reliable enough to build workflows around.

From what we have seen working with frontier models on complex client problems, the answer is: sometimes, for some problem types. Models at the GPT-5 level are genuinely better at holding a long chain of domain-specific reasoning together than earlier generations. But they still hallucinate, still miss domain nuance, and still need an expert in the loop to catch errors. Unutmaz is an immunologist. He would know if GPT-5 gave him a plausible-sounding but wrong answer. Most users are not in that position.

The practical pattern worth noting is this: the scientists getting the most out of these models are not using them as search engines. They are using them as a thinking partner for a specific, well-scoped problem they already understand deeply. That is a different skill than prompting for a summary.

What to do about it

If you are doing any kind of structured research or analysis in your business, whether that is competitive research, customer data interpretation, or technical problem-solving, try scoping a problem you have been stuck on and working through it with GPT-5 Pro or a comparable model. Do not ask for a general overview. Bring your specific data, your specific question, and push the model to reason through it step by step. Then have a domain expert check the output before you act on it.

The researchers getting real value from these tools are the ones who treat the model as a rigorous collaborator, not a fast answer machine.

Source: OpenAI Blog