AI models have performed well on mathematics benchmarks before. OpenAI’s latest claim goes further, placing an unreleased model inside active research problems where the answers were not already known.
OpenAI says an internal version of Astra produced 10 advances in mathematics and theoretical computer science. The company carefully describes the results as either resolving a longstanding problem or making substantial progress, rather than claiming that every problem was completely solved.
The work covers high-dimensional geometry, coding theory, group theory, operator algebras, arithmetic circuit complexity, quantum complexity, lattice cryptography and extremal combinatorics. The complete collection was published as a 249-page manuscript.
Among the most significant results is the construction of an explicit non-sofic group, addressing a major question that had remained open since Mikhail Gromov introduced the concept in 1999. Astra also produced a disproof of Connes’s rigidity conjecture and results resolving three problems from Paul Erdős’s catalogue, numbered 146, 180 and 183.
Other advances include new bounds for sphere packing and error-correcting codes, lower bounds in arithmetic circuit complexity, a theorem covering two-player quantum games and progress on the closest vector problem, which is connected to research in post-quantum cryptography.
The verification process is as important as the results. OpenAI says the model generated the mathematical arguments, after which humans prepared them as manuscripts with assistance from Astra. The model then formalized each argument as a Lean certificate, allowing proof-checking software to verify whether the logical steps follow correctly.
OpenAI estimates that the model tokens used to find the results would have cost approximately $2,000 at Sol API rates. That figure covers estimated inference usage only and should not be treated as the total cost of human review, formalization or research preparation.
The announcement has also reopened questions about credit and recognition in AI-assisted research. OpenAI argues that attribution should reflect how a result was produced and says it would be misleading to present an AI-generated proof as entirely human work. The Economic Times also reported debate over whether traditional honours can adequately recognize research created through a combination of models, engineers and mathematicians.
One point needs careful separation. Fields Medalist Tim Gowers’ positive assessment concerned OpenAI’s earlier AI-generated unit-distance proof, not all 10 Astra results as a group. That earlier work helped establish that AI-generated mathematics could reach a standard worthy of serious expert review.
Astra’s importance therefore rests on more than producing impressive answers. The stronger claim is that an AI system generated new mathematical arguments, converted them into formally checkable proofs, and created results that researchers can now examine, challenge and extend.
Source : Economic Times