OpenAI says an internal model called Astra produced new results on 10 longstanding problems spanning mathematics and theoretical computer science, according to the company's own account. No paper, preprint, or third-party lab has verified any of the 10 claimed results, and OpenAI has not named the problems, the venues that would review them, or a date for independent replication.

The pattern is familiar from DeepMind's FunSearch and AlphaEvolve work: a language model searches a space of proofs or constructions and a company announces the count before peer review catches up. The next gate is simple and hasn't been cleared: a mathematician outside OpenAI, working from a published proof rather than a press account, checking one of these 10 results line by line.