OpenAI is slowing down some of its own AI training, and it's doing so on purpose.
On Tuesday, the company said it had paused reinforcement learning training on the latest models it plans to deploy, a freeze expected to last about two weeks. It's also delaying what it calls its largest planned frontier RL run, with no set date given for resuming. OpenAI framed the move as time to tighten security and safeguards before pushing ahead. The company did not specify what prompted the review or when the delayed run will restart.
This is notable mainly for what OpenAI is choosing not to do: race. The company is heading toward a looming IPO and facing aggressive competition from Anthropic plus fast-improving open-weight and Chinese models, all of which reward speed over caution. Pausing anyway is a real-world test of an argument AI safety researchers have made for years, that a leading lab should be willing to voluntarily slow down rather than assume competitors will do the same.
Whether this marks a genuine shift in posture or just a pause before the next frontier run resumes at full speed is the part worth watching.