OpenAI prices frontier monitoring at 20% compute

OpenAI says its expanded monitoring for tool-using reinforcement learning and high-risk inference carries an estimated 20% compute overhead and aims to alert teams within 30 minutes of concerning activity. In the company’s new safeguards post, OpenAI says the policy follows a two-week RL pause and keeps its largest planned frontier run on hold while it validates alignment and containment; its incident account describes the earlier evaluation breach at Hugging Face. Agent teams should budget monitoring as a first-class inference line item rather than a zero-cost wrapper, a governance cost that complements the archive’s evidence on agent evaluation scope.