A new report from OpenAI and a group of outside scientists shows how GPT-5, the company’s latest AI large language model (LLM), can help with research from black holes to cancer‑fighting cells to math puzzles.
Each chapter in the paper offers case studies: a mathematician or a physicist stuck in a quandary, a doctor trying to confirm a lab result. They all ask GPT-5 for help. Sometimes the LLM gets things wrong. Sometimes it finds a faster route to an already known result. But other times, with careful human guidance, it helps push the boundaries of what was previously known.
In one experiment involving how waves behave around black holes, GPT-5 worked through the math to independently produce results that had previously been shown to be correct, showing it was capable of doing this level of scientific calculation. In another project involving nuclear fusion, GPT-5 developed a model that accelerated the research. “AI’s ability to dramatically reduce the time required for coding—compressing what would traditionally take days into mere minutes for the author—has monumental implications for research practices,” says Floor Broekgaarden, an astronomer at the University of California, San Diego, who was not involved in the study.
On supporting science journalism
If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
In another case, researchers studying immune cells used GPT-5 to interpret their data, and its explanation matched results the lab had already confirmed. “GPT-5 Pro can function as a true mechanistic co-investigator in biomedical research, compressing months of reasoning into minutes, uncovering non-obvious hypotheses, and directly shaping experimentally testable strategies,” Derya Unutmaz, the doctor leading the project, wrote in the paper.
The paper also announces several new math discoveries supported by GPT-5. Guided by human experts, it solved a long-standing problem posed in 1992 by mathematician Paul Erdős. It also produced a clearer rule showing the limitations of how computer systems make decisions; discovered another rule for how certain small patterns appear inside branching diagrams; and found a way to spot secret structures in a network as it grows. The discoveries are modest but appear to be genuine, and each was verified by human mathematicians.
“I had not seen anything that impressive [in math] from an LLM before,” says Ryan Foley, an astrophysicist at the University of California, Santa Cruz, who was not involved in the study. “I suspect LLMs are going to upend how theories are created, vetted and improved.” He cautions, however, that AI tools still require significant prompting: “Humans are creative; AI is responsive. However, the rate of discovery should rapidly increase.”
Prithviraj Ammanabrolu, a computer scientist at the University of California, San Diego, who was not involved in the research, points out that the published work is more a series of case studies than a scientific paper because it doesn’t provide enough details to repeat the experiments and doesn’t offer counterfactual analysis involving different approaches. Despite these limitations, AI’s ability to help with research “is still miles ahead of what was possible even a year ago, so the rate of progress is quite high,” he says. “It shows future potential in enabling scientists to accurately mix together relevant prior results and draw new insights in novel ways.”
One of GPT-5’s strengths is its ability to search vast quantities of scientific literature. For a math problem listed as unsolved online, it identified a solution in a paper from the 1980s. In another case, it found a few lines in a German paper from the 1960s that settled a problem. It easily navigated the language barrier and the differences in style between midcentury math writing and contemporary approaches.
All of this might make GPT‑5 sound like a scientific genius, but the paper’s authors are clear that it’s not. Rather, in the right hands, it is a fast and tireless assistant that has read an impossible number of papers and never minds reworking a calculation. But human judgment is not optional, they stress. Researchers also caught it being confidently wrong, and it can misstate references, hallucinating nonexistent papers or failing to credit authors of real ones.
“Human expertise remains crucial,” Broekgaarden says. But AI “can take on myriad tasks—collating data, summarizing research articles, and even performing complex calculations—that previously demanded extensive time and effort from researchers.”
Numerous ways that AI will shape research remain to be seen. New AI models are released every few months. If general‑purpose chatbots that struggled with middle school math two years ago can now spot hidden structures in black-hole waves and suggest new approaches to cell therapy, who knows what their successors will achieve?