OpenAI says its unreleased “Astra” model family can crack ten previously difficult math problems, demonstrated in a mathematics paper released alongside supporting materials. According to multiple reports, the company frames the announcement around a technical evaluation rather than a traditional product launch. For each of the ten results, OpenAI provides machine-checkable proofs written in Lean 4 and hosted on GitHub. The outlets also report that the effort to produce the results is described as costing roughly $2,000 in computing resources. One of the highlighted achievements includes an explicit construction related to a non-sofic group, presented as part of the set of solved problems. The proofs are intended to be independently verifiable via the Lean 4 formalization, allowing readers to check the correctness of the claimed results rather than relying solely on narrative descriptions. OpenAI has not broadly announced availability or release timing for Astra, but the company’s publication emphasizes the model’s capabilities through the formal proof artifacts.