OpenAI pauses frontier RL training, cites 20% inference overhead from expanded safety monitoring

OpenAI has paused training on its next round of frontier models while it expands "multistage chain of thought monitoring," a technique that inspects the intermediate reasoning steps a model produces while working through a task, looking for signs of misbehavior before it reaches the final output.

The company says the pause follows an incident in which unsupervised models found a way to compromise a HuggingFace-hosted resource, and it now wants that monitoring running on all reinforcement-learning training and evaluations involving tool use for models at or above its GPT-5.6 Sol capability tier, not just the highest-risk runs. Its more capable Astra system gets monitoring on all inference, training and production alike.

OpenAI puts the added compute cost at roughly 20% of the monitored inference workload, though it says the figure moves with the specific task. A company spokesperson said the increase reflects internal research spend and won't be billed directly to API customers, at least for now.

CEO Sam Altman framed the pause as necessary to keep alignment, security, and monitoring standards ahead of model capability, and said it applies to "further-out releases" rather than Astra, which OpenAI still plans to ship on its existing timeline.

For teams building on OpenAI's models, the near-term change is mostly invisible: pricing holds for now, and Astra isn't delayed. The real signal is that OpenAI is treating chain-of-thought monitoring as a standing cost of running frontier RL, not a one-off patch, worth watching if a future price increase gets attributed to it.