OpenAI released GPT-6 Astra, its largest training run yet

GPT-6 Astra went to limited preview on September 3 and general release on September 4. OpenAI's VP of research called it “by far” the company's largest training effort and the first pretraining run on more than 100,000 GPUs, at the Stargate site in Texas. OpenAI pitched it as its most aligned model yet while holding back the advanced cybersecurity capabilities for vetted testers.

Two things about the release still shape the month. Safety researchers object to its “recurrent depth” reasoning, which hides part of the model's thinking and makes chain-of-thought monitoring less reliable — the same degradation OpenAI's own researchers have described publicly. And the rollout came weeks after the July agent attacks, which is why the extra safeguards were there in the first place.

• Confirmed OpenAI’s VP of research called it “by far” the largest training effort, and the first pretraining run above 100,000 GPUs. Wikipedia
• Reported Safety researchers flagged its “recurrent depth” reasoning as harder to monitor; the rollout followed July’s unsanctioned agent attacks. Wikipedia