In a recent research paper, the company revealed that GPT-5 helped a mathematician from Columbia University find a solution to a famous and complex equation known as the “Erdős” problem in number theory, a problem that had remained unsolved for many years.
According to the company, the model also succeeded within minutes in identifying a critical change within human immune cells, a discovery that took scientists months of research. In addition, the model suggested a laboratory experiment that researchers were able to implement and verify.
Kevin Weil, Vice President of Science at OpenAI, said that putting these tools within reach of researchers around the world could enable the completion of “twenty-five years of scientific research in just five years.”
These developments come as technology companies are strongly entering the scientific sector. Last month, Anthropic announced the integration of its “Claude” model into the tools used by researchers and life science companies. At the same time, Google unveiled the “Partner Scientist” tool, which helps formulate hypotheses, and confirmed that the open-source “Gemma” model contributed to the discovery of a potential new pathway for cancer treatment.
In a move to enhance its scientific ambitions, OpenAI created a new science unit in October and appointed theoretical physicist Alex Lopsachka, known for his research on black holes, as chief researcher. The company also plans to build a fully automated scientific research system by March 2028.
Specialized researchers confirm that “artificial intelligence is already capable of driving scientific discoveries,” noting that these models have not yet reached the level of “independent robotic scientist,” but rather act as a “research assistant” with broad access to scientific papers and quantitative tools, but only when guided by a human expert.
OpenAI confirmed that GPT-5 is particularly adept at in-depth research within scientific literature, linking knowledge distributed across multiple languages and fields. The model also rediscovered new research results that were not part of its training data, which was considered a powerful example of its ability to “extract what was hidden from the human eye.”
The research paper published by OpenAI clarifies that researchers relied on GPT-5 to rediscover old research on the “Erdős” problem that was buried within specialized journals, German margins, and lengthy reviews that modern studies had not noticed, which is considered one of the model’s most prominent aspects.
At the same time, the paper also emphasized that the model is still prone to hallucinations and the fabrication of false information, necessitating human scrutiny in formulating the problem and correcting hypotheses and outputs.
Even with its tangible contributions, the role of the GPT-5 model remains an assistant, not a leader. In almost all cases, scientists were the ones who identified the problem, developed the strategy, and made the fundamental decisions. The model provided auxiliary materials such as proof schemes, digital experiments, and hypothesis formulations. Researchers agree that the essence of discoveries remains human, but artificial intelligence has become a tool capable of opening new paths that were not easily accessible before.
