1 OpenAI put a price on the compute required to monitor AI
On Aug. 18, OpenAI disclosed that monitoring models during reinforcement learning training and evaluation can consume roughly 20% of the inference compute being monitored. The company also said it had paused reinforcement learning training for two weeks while it evaluated safety controls for planned frontier models.
- Aug. 18 — OpenAI said monitoring is required for reinforcement learning training and tool-using evaluations involving models at Sol capability or above.
- Aug. 18 — The company estimated monitoring overhead at roughly 20% of the inference compute being monitored, while noting that the cost varies across workloads.
- Aug. 18 — OpenAI said its largest planned frontier training run remained on hold pending smaller-scale evaluations.
- Aug. 7 — After determining that it could not rule out critical cyber capability in Astra, OpenAI extended monitoring to all Astra inference using tools. The company said the monitoring is designed to alert within 30 minutes of concerning activity.
Monitoring has usually been discussed as a security control. OpenAI has now put a measurable compute cost on running it. That changes the operating conversation: monitoring is part of the infrastructure required to run some AI workloads, not just a policy decision made around them.
The pause is significant too. OpenAI is showing that stronger safety controls can affect when and how a model is trained and released, not just what happens after deployment.
Operator takeaway: Ask providers what monitoring runs alongside your production workloads, how much compute it consumes and how that overhead is reflected in your capacity and cost planning.