OpenAI recently announced that its internal models have solved ten long-standing problems in mathematics and theoretical computer science. These results, which address questions that had remained open for at least a decade, span fields such as high-dimensional geometry, group theory, coding theory, and quantum complexity. Among the notable achievements is a construction establishing the existence of non-sofic groups and new bounds on high-dimensional sphere packing.
The solutions were generated by an internal version of the model known as Astra. According to OpenAI, the process involved human collaboration to prepare manuscripts, while the model formalized each proof into a Lean certificate. This approach highlights a shift in how research is conducted, as the total token cost for these solutions reached approximately $2,000 at current rates.
OpenAI stated that it intends to provide 100,000 scientists and mathematicians with free access to its models to accelerate discovery. The organization acknowledged the concerns regarding the role of automated systems in academic research. They maintain that claiming human authorship for entirely machine-generated proofs would be a misrepresentation of intellectual work. Instead, they frame these results as contributions intended for the broader scientific community to verify, debate, and build upon.
The company emphasized that these models serve as collaborators rather than replacements for human inquiry. By releasing both the proof and the model’s narration of its thinking process, OpenAI expects researchers to test the validity of these outcomes. This initiative marks a significant step in determining how computational power can address foundational challenges in mathematics, while addressing ethical considerations regarding authorship and transparency in peer-reviewed science.