On August 19, OpenAI announced unscheduled news through its official account: it had paused part of the training of its latest model for two weeks.
OpenAI Pauses Frontier Training for Two Weeks
According to the official post, it temporarily paused reinforcement learning (RL) training of its latest model intended for deployment for two weeks as of August 19, 2026 (KST). The purpose is to harden its research environment, run red-team verification, and widen monitoring. The largest frontier RL training is still on hold.
Co-founder Greg Brockman also quoted the announcement the same day and added that "confidence in safety will increasingly set the pace of AI development."
Has Astra Crossed a Dangerous Line?
OpenAI's official blog post "Pacing model development in an era of cyber-critical capabilities" explained the background to this decision. Preliminary evidence has emerged that its next-generation model, "Astra," may have reached the "Critical" cybersecurity capability threshold under its own Preparedness Framework.
The Critical level means a level at which a model could find zero-day vulnerabilities in hardened real-world systems and produce exploits without human involvement. If, given only a high-level goal, it could devise and carry out novel attack strategies on its own, it would also fall into this tier. This is the first time OpenAI has publicly mentioned the possibility of reaching this threshold.
The Next Model's Release May Slip
Some workloads directly tied to safety and alignment have been moved first to the new security standards. A good share of the remaining workloads stay paused until migration and hardening are finished.
If your team builds products on the OpenAI API or GPT-family models, keep in mind that new frontier model release schedules could slip for a while. At the same time, this step is also a sign that the trend of security verification procedures setting the pace of deployment as model capabilities grow has begun in earnest.
Paused or Slowed: A Difference in Words
The official @OpenAI post used the word "paused." But in the post Brockman quoted, the wording changed to "slowed." In replies to that post, one user commented on the word difference: "Going by this wording alone, it could be total safety theatre or it might not be. It's impossible to tell." That is only one user's interpretation, and OpenAI hasn't officially explained the difference between the two words yet.
Whether Astra has actually reached the Critical level, or is only at the preliminary-evidence stage, also remains in "not officially confirmed" territory. OpenAI said only that it is verifying safeguards through small-scale training and evaluations. The exact restart date and final evaluation results will have to wait for follow-up announcements.