OpenAI paused reinforcement-learning training on its largest planned frontier run for two weeks after determining, on August 7, that its unreleased Astra model may cross the Critical cyber threshold under the company's Preparedness Framework. The disclosure came in an August 18 post titled "Pacing model development in an era of cyber-critical capabilities," and it's the first time OpenAI has publicly slowed a training run on internally derived safety grounds since the framework was first published in 2023.

The tier distinction is where the news lives. Under the Preparedness Framework, a High rating triggers pre-deployment safeguards; a Critical rating requires halting internal development. GPT-5.6 Sol, and the model implicated in the July 2026 Hugging Face incident, were both assessed at High. Per Forbes, Astra is the first OpenAI model to breach the Critical line, and no outside body has independently verified that classification. Axios reports the framework itself is now being rewritten.

The operational disclosures are unusually specific. Chain-of-thought monitoring is now required across all RL training and evaluations involving tools for models at Sol capability or higher, and, after August 7, on all Astra inference with tools, not just training. Activation classifiers inspect the model's internal state at every sampled token and escalate to automated investigators. A likely violation of a critical security boundary pages the safety, security, and research teams. Alerts must surface within 30 minutes, and those teams have the same 30-minute window to pause activity if they can't conclusively rule out a false positive.

None of this is free. "Our current estimates put monitoring overhead at roughly 20% of the inference compute being monitored, though the cost varies substantially across training and evaluation workloads." An OpenAI spokesperson told The Register that overhead won't be passed to customers, which is itself a strategic tell: OpenAI is choosing to absorb a five-point margin haircut on some workloads rather than reprice safety as a line item.

The political staging matches the technical staging. A White House official told Axios that OpenAI "voluntarily informed the administration of their plans to delay the release." Sam Altman said on social media that near-term models, presumably Astra, will still ship, and that the pause affects further-out releases. Read together, this is narrative management aimed at Washington: proof of restraint without a product delay anyone will feel in Q4.

The industry contrast lands harder. Per Axios, "Anthropic previously committed to pausing training of powerful models if capabilities surpassed the company's ability to control them. But the AI lab rolled that back in an update to its Responsible Scaling Policy in February of this year." OpenAI just filled the vacancy Anthropic vacated six months ago, and priced the seat at 20% of inference compute.

Sources

https://openai.com/index/pacing-model-development-cyber-capabilities/
https://www.theregister.com/ai-and-ml/2026/08/19/openais-overhead-will-rise-20-percent-for-some-workloads-as-it-hardens-security/5289303
https://www.techtimes.com/articles/324929/20260819/openai-hacked-hugging-face-then-deployed-safety-monitors-its-own-scientists-proved-can-gamed.htm
https://www.axios.com/2026/08/18/openai-pause-astra-preparedness-framework
https://www.forbes.com/sites/ashishbhatia/2026/08/19/openai-paused-ai-training-for-two-weeks-heres-what-that-means/