Separating the confirmed Astra facts from the unsourced ones: published proofs, a cyber tier, a training pause

OpenAI has stated that an internal version of Astra produced results resolving or substantially advancing ten longstanding problems in math and theoretical computer science, with a 250-page manuscript and machine-checkable certificates. OpenAI noted caveats, including in-house problem selection and human-prepared write-ups. The company also cannot rule out that the model meets its highest cyber tier preparedness framework and paused reinforcement learning for deployment-bound models for two weeks, increasing monitoring costs by approximately one-fifth. No release date or naming decision has been announced.