OpenAI reports sharply lower reasoning-extraction success on GPT-6.2 Astra and describes campaign-level enforcement.
A 99-problem benchmark drawn from 117 papers and patents; GPT-6.2 Astra scores 84.5%.
OpenAI describes a cloud system of Codex-powered workflows that root-causes and proposes fixes for performance regressions.
Model-proposed changes cut standard-cell area by about 10-13% in Jalapeno, and AMD reports up to 14% PPA gains; OpenAI plans GPT-Synopsys with Synopsys.
Codex uses native OS sandboxing with permission profiles for filesystem and network limits, and is moving its Windows sandbox to Microsoft's MXC.
OpenAI outlines four defense layers across research, internal deployment and external deployment, and discloses a 2026 containment incident.
A follow-up report finds autonomous task success rising and agent usage growing sharply, while uplift remains uneven.
Astra for Law ranked first on Harvey LAB-AA; ChatGPT Enterprise customers can request access through the Trusted Access Program.
A new Sales plugin plus more than 30 new and expanded partner plugins cover prospecting, CRM, meetings, quotes and agreements in ChatGPT.