OpenAI safety monitoring adds ~20% inference compute for GPT-5.6 Sol and above
OpenAI on Aug 19 disclosed that hardened chain-of-thought monitoring adds ~20% inference compute, covering all RL training and tool-involving evaluations for GPT-5.6 Sol-class and above, plus all Astra inference.
In its Aug 19 disclosure on the training pause, OpenAI detailed the compute cost of its hardened safety monitoring. The new multi-stage chain-of-thought monitoring covers all reinforcement learning training and tool-involving evaluations for GPT-5.6 Sol-class models and above, and applies to all inference on Astra.The monitoring aims to flag anomalous behavior during chain-of-thought reasoning, with the goal of issuing alerts within 30 minutes. OpenAI expects the overhead to add roughly 20% to inference compute for monitored workloads.On Aug 18 the company paused reinforcement learning training on its latest deployment-track models for two weeks. Its largest planned frontier RL run remains on hold while the company runs smaller-scale training and evaluations to validate safeguards. Lower-risk training has resumed.OpenAI disclosed that after the July Hugging Face incident, it added tighter sandbox isolation, network controls and multi-stage monitoring. An internal evaluation on Aug 7 found that Astra may have reached the Preparedness Framework Critical cybersecurity tier.