OpenAI paused frontier RL training for two weeks after evidence its coming Astra model may hit a Critical cyber threshold, holding its largest run behind token-level classifiers and a 20% compute toll. Sam Altman: confidence in safety will increasingly set the pace of AI progress.
The caveat that matters: Mythos 2 is trained and withheld while the loop building Mythos 3 runs on. What is paused is the export, not the engine.