OpenAI has announced that its internal model, Astra, has successfully provided solutions to ten long-standing mathematical problems, some of which have remained unsolved for decades. The achievement, detailed in a 249-page manuscript, spans diverse fields including group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography, and extremal combinatorics. This development marks a significant moment for artificial intelligence in scientific research, particularly in mathematics, a field traditionally reliant on human intellect and rigorous proof.
A key aspect of Astra's contribution is the provision of machine-checkable certificates for each result. This feature allows for instant, trustless verification of the proofs, bypassing the lengthy and often subjective traditional peer-review process. The problems addressed were not trivial exercises but central questions within their respective subfields. Among the solved problems is a construction establishing the existence of non-sofic groups, a question that has occupied group theorists for years, and a disproof of Connes's rigidity conjecture in the theory of von Neumann algebras. Additionally, Astra improved the general upper bound on sphere-packing density in high dimensions, a bound that had been in place since 1978. Three of the problems originated from the catalogue of open questions left by the prolific mathematician Paul Erdős.
This latest announcement follows a similar breakthrough in May 2026, when a related OpenAI model reportedly disproved the Erdős unit distance conjecture, an 80-year-old problem in discrete geometry. The proof for that conjecture was verified by a team of external mathematicians, who also published a companion paper to help human researchers understand the AI-generated argument. The fact that a general-purpose reasoning model, rather than one specifically trained for mathematics, produced these proofs is seen as particularly noteworthy.
The implications of AI's increasing capability in mathematical research are profound and have ignited debate within the mathematical community. Some mathematicians express concern about the potential existential crisis AI could pose to the field, questioning the future role of human mathematicians if AI can generate novel proofs and solve complex problems. The Leiden Declaration on Artificial Intelligence and Mathematics highlights that unchecked trends in AI could threaten researchers' autonomy and affect the scope and depth of mathematical research itself.
However, others view AI as a powerful tool that can accelerate discovery and aid human researchers. AI methods are already helping scientists explore complex systems, analyze large datasets, and generate new hypotheses across various disciplines. The interaction between AI and mathematical sciences is seen as a mutually beneficial relationship that has the potential to speed up progress in both areas. The development of AI tools capable of performing research-level mathematical tasks necessitates a re-examination of the goals and values of mathematical research, with problem-solving being a key area of focus.
While the cost of producing these ten solutions was reportedly only $2,000, raising questions about the economics of mathematical discovery, the accessibility and perceived bias in problem selection are also points of discussion among critics. Nevertheless, the verifiable nature of Astra's output addresses a significant challenge in the deployment of AI in scientific fields, offering a new paradigm for mathematical validation.