GPT-6 Astra is positioned by OpenAI as its next flagship model, stepping into a tier of capabilities that blends deep reasoning with serious cybersecurity performance. Reported evaluations place it at the Critical level on the Preparedness Framework's cyber category, a first for an OpenAI release, alongside a reported 100% score on ExploitBench and a 91.5% refusal rate on cyber jailbreaks, a figure OpenAI contrasts against 59% for GPT-5.6 Sol. The model is also credited with finding two chained zero-days on an internal V8 test set, signaling a shift toward frontier offensive-security research as part of its stated capabilities.
Beyond cybersecurity, GPT-6 Astra is associated with formal mathematical progress. An August 2026 OpenAI report describes the model solving ten previously open problems in mathematics and theoretical computer science, with proofs formalized in Lean, indicating strong tool-assisted reasoning over rigorous symbolic domains. Coverage also describes a recurrent-style architecture that reuses the same layer set in loops to add compute per token, a design choice that aims to scale reasoning depth without proportionally expanding parameter count. Together, these traits frame GPT-6 Astra as a model built for sustained, multi-step problem solving across both formal reasoning and high-stakes technical tasks.
