The new batch of results includes a solution to the four-dimensional Kakeya conjecture, improvements to major computer algorithms, and actual progress toward the Riemann hypothesis. The release comprises 722 manuscripts covering 372 result families. According to OpenAI, almost every result was generated in response to a single prompt handed to a single AI agent. This represents a significant methodological shift from the company’s earlier solution to the Navier-Stokes problem, which required the collective efforts of a 10,000-strong agentic swarm and incurred millions of dollars in computing costs. OpenAI states that the average result from the new batch utilized the equivalent of three hours of ChatGPT Pro thinking time.
Despite the volume of new output, significant skepticism remains regarding the transparency of the process. An independent advisory group of elite mathematicians, recently assembled by OpenAI, recommended that any company releasing AI-generated results should publicly disclose the model, the exact prompts used, and the compute time for each proof. In contrast, OpenAI has chosen to reveal only the average compute time and general statistics, withholding the specific prompts and keeping the model proprietary. The company stated it is not bound by the advisory group’s recommendations but claims to be taking the guidelines seriously.
Many of the released results have already been verified in Lean, a programming language designed to validate the logic of mathematical proofs. This verification suggests a high degree of correctness. However, the broader mathematical community remains divided on the implications of such rapid, unexplained advances. Critics argue that the lack of access to the underlying model prevents independent replication and full understanding of whether the proofs contain novel ideas or are merely combinations of existing techniques.
Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology, expressed doubt about the company’s claims regarding the efficiency of the process. He noted that until the model is released and results can be replicated, claims about solving problems with a single agent should be treated as unverified. The tension is further complicated by OpenAI’s previous handling of the Navier-Stokes solution, which was criticized for a lack of clarity and adherence to academic norms. The advisory group was formed specifically to address these concerns and provide recommendations on the responsible publication of AI-generated mathematical research.
The release highlights a growing friction between AI developers and the academic community. While OpenAI claims to be working to release the model as quickly and responsibly as possible, the current approach of withholding proprietary details has drawn criticism. Mathematicians worry that the “interloper” nature of these well-resourced AI systems, which operate with different incentives than traditional researchers, may disrupt long-standing norms of collaboration and attribution in the field.
The immediate practical effect of this release is a surge in unverified yet logically sound mathematical content. The next step for OpenAI remains the potential public release of the model itself, a milestone that could fundamentally change how mathematical research is conducted and verified. Until then, the community faces the task of parsing hundreds of new proofs while waiting for greater transparency from the company.
