OpenAI says Astra produced 10 new math and TCS results
Original: Ten advances in mathematics and theoretical computer science View original →
AI-assisted science just moved into more contentious territory: not only helping researchers code or search literature, but generating candidate mathematical results on problems that have been open for years. OpenAI published a set of 10 advances on August 1, 2026, spanning mathematics and theoretical computer science.
The concrete claim is unusually large. OpenAI says the results were produced by an internal version of Astra, its next major model, and that the total token usage required to find the solutions would cost roughly $2,000 at Sol API rates. Humans then prepared the arguments into manuscripts with the same model, after which the model formalized each argument into a Lean certificate.
The list covers high-dimensional sphere packing, binary and spherical codes, non-sofic groups, Connes rigidity, arithmetic circuit complexity, quantum parallel repetition, the closest vector problem in lattice cryptography, Ehrhart volume, multicolor Ramsey numbers, and extremal graph theory. Several entries touch problems with broad implications across mathematics, complexity theory, and cryptography.
The Lean layer is the part to watch. Mathematical AI claims are easy to overstate when they are only prose. By releasing formal certificates and reasoning walkthroughs, OpenAI is giving specialists something more concrete to inspect: not just a polished write-up, but a route for independent verification, correction, or rejection.
The caveat is equally important. A company posting manuscripts is not the same as the mathematical community accepting the results. OpenAI says it takes responsibility for correctness while also acknowledging the attribution problem: if a proof was generated by an AI system, presenting it as ordinary human authorship would misrepresent how the work was produced. The next phase belongs to independent mathematicians reading the certificates, testing the arguments, and deciding which pieces hold up.
Related Articles
OpenAI’s next major model family, Astra, is being tested through research outputs rather than only benchmarks. The company says an internal version produced 10 results and that finding them would cost roughly $2,000 at Sol API rates.
OpenAI says ChatGPT is already being used at research scale across science and mathematics. In its January 2026 report, the company says advanced science and math usage reached nearly 8.4 million weekly messages from roughly 1.3 million weekly users, with early evidence that GPT-5.2 is contributing to serious mathematical work.
HN read this math story less as another "AI did it" headline and more as a case where a model pointed at a route humans had not tried. The part that stuck was the expert cleanup work after the GPT-5.4 Pro draft, not the one-shot prompt itself.