A recent report from OpenAI, in collaboration with outside scientists, has demonstrated the remarkable capabilities of GPT-5, the company’s latest large language model (LLM), in assisting scientific research across diverse fields such as astrophysics, mathematics, cancer biology, and nuclear fusion. The findings, published as a collection of detailed case studies, showcase how GPT-5 is emerging as a powerful tool that helps researchers test hypotheses, navigate vast scientific literature, and refine experimental approaches—marking a significant step toward integrating AI into the scientific process.
One of the standout examples in the report involves the study of waves around black holes. In this case, GPT-5 independently worked through complex mathematical calculations and produced results consistent with previously validated findings in theoretical physics. This demonstrated the AI’s capacity not only for understanding sophisticated scientific concepts but also for performing high-level scientific calculations with accuracy. Similarly, in a project focused on nuclear fusion, GPT-5 developed a computational model that accelerated the research timeline drastically. Floor Broekgaarden, an astronomer at the University of California, San Diego, noted the profound implications of this ability, emphasizing that GPT-5’s capacity to reduce coding time from days to minutes could revolutionize how scientific research is conducted.
Beyond physics, GPT-5 also showed impressive contributions in biomedical research. In one study on immune cells, researchers used GPT-5 to analyze experimental data and generate interpretations that perfectly aligned with the team’s confirmed laboratory results. The lead researcher, Dr. Derya Unutmaz, described GPT-5 Pro as functioning like a “true mechanistic co-investigator,” capable of compressing months of complex reasoning into mere minutes, proposing non-obvious hypotheses, and shaping experimental designs that can be directly tested in the lab. This highlights the AI’s potential to not only assist with data analysis but to actively contribute to scientific discovery and experimental strategy.
Mathematics, a field traditionally resistant to automation, also benefited from GPT-5’s abilities. Guided by human mathematicians, the AI helped solve a longstanding problem originally posed in 1992 by the renowned mathematician Paul Erdős. Additionally, GPT-5 uncovered new mathematical rules concerning the limitations of computer decision-making, identified patterns in branching diagrams, and devised methods to detect hidden structures within growing networks. Although these discoveries are modest in scale, their validation by human experts confirms that GPT-5 can contribute meaningfully to mathematical research. Ryan Foley, an astrophysicist not involved in the study, remarked that such mathematical accomplishments by an LLM are unprecedented and suggested that AI could fundamentally transform how scientific theories are developed, evaluated, and refined. However, he cautioned that AI remains a responsive tool that requires human creativity and guidance.
Despite the excitement surrounding GPT-5’s capabilities, some experts urge caution. Prithviraj Ammanabrolu, a computer scientist at UC San Diego, noted that the published work reads more like a series of case studies than a traditional scientific paper, as it lacks the detailed methodologies necessary for full reproducibility and does not explore alternative approaches or counterfactuals. Nonetheless, he emphasized that the rapid progress in AI-assisted research is undeniable, with GPT-5’s ability to synthesize prior results and generate new insights representing a significant leap beyond what was possible just a year ago.
One key strength of GPT-5 lies in its ability to rapidly search and analyze vast amounts of scientific literature, including sources spanning decades and multiple languages. For example, when presented with a mathematical problem considered unsolved, GPT-5 identified a solution documented in a paper from the 1980s. In another instance, it located a critical passage in a German-language paper published in the 1960s, resolving an open question. The AI’s facility in overcoming language barriers and adapting to historical differences in scientific writing style enables it to bring otherwise overlooked knowledge to the forefront of current research.
While GPT-5’s performance might sometimes suggest the presence of a scientific genius, the authors of the report are quick to clarify that the AI is not a substitute for human researchers. Instead, it functions as an extraordinarily fast and tireless assistant, capable of digesting an immense volume of scientific papers and repeatedly running complex calculations without fatigue. Crucially, human expertise remains indispensable. Researchers observed instances where GPT-5 was confidently incorrect, misrepresented references, fabricated nonexistent studies, or failed to properly credit real authors. This underscores the importance of careful human oversight and critical evaluation when using AI tools in research