OpenAI, Harvard, Oxford and other top university paper shows GPT-5 already helps real scientists push forward actual research in several scientific fields across math, physics, biology, and more.
In several projects, GPT-5 helped prove 4 new math results and checked tricky steps that humans then fully verified.
GPT-5 has started actually helping solve fresh research problems.
The authors set up many small case studies where experts dropped GPT-5 into live work in mathematics, physics, astronomy, computer science, biology, and materials science.
In each case, GPT-5 suggested research ideas, filled missing proof steps, wrote and debugged code, searched the literature, and pointed out gaps in arguments.
Some examples are tightening a rule about safe step sizes in an optimization method, finding hidden symmetries in black hole equations, and explaining puzzling immune cell experiments.
Across these studies, GPT-5 is very good at step by step reasoning and exploring many options quickly, but it still makes real mistakes and cannot reliably judge if a result is correct.
Overall, the paper treats GPT-5 as a strong junior collaborator that saves expert time while humans still design the plan and carefully check every important step.
----
Paper β arxiv. org/abs/2511.16072
Paper Title: "Early science acceleration experiments with GPT-5"
βKorean people have an indomitable will. Although there are many difficulties in front of us, I believe that we will be able to prevail.β South Koreaβs President sits down with TIME in his only Western media interview since taking office https://t.co/zVRLgiEKt2