OpenAI says GPT-6.2 Astra and GPT-6.2 Sol find and exploit flaws better and write more secure code. The obvious reading is defenders just got faster help.
I am skeptical of that payoff. OpenAI also rates both models at its Critical threshold for cybersecurity.
That rating triggered tighter safeguards for higher-risk help, OpenAI said. The same filters will restrict some legit vulnerability research and testing. Approved defenders keep reduced refusals on Daybreak Blue and GPT-6 Cyber on Daybreak Red, while GPT-6.2 waits for Daybreak access.
OpenAI tested cyber skills with safeguards disabled, per its system card. It names DeepBinary, ExploitBench, SecureSWE and PatchBench but gives no exploit rates in the announcement. For the obvious reading to hold, the patches must preserve behavior at scale and secure-code gains must survive real repos.
To be fair, OpenAI pairs capability with efficiency. It prices Astra at $8 input, $0.40 cached input and $40 output per million tokens, 20% below GPT-6 Astra, per OpenAI. Astra for Law averaged 0.03 material errors per assignment, half the base rate and below Claude Opus 5.5, per OpenAI.
Both models also hit High for biological and chemical capabilities, OpenAI said. Astra is the first external release to hit High for AI self-improvement. If you lead a security team, pilot GPT-6.2 on your own backlog before you change staffing.





