OpenAI says its new Control Room system halts model output within 35 seconds of a decision to intervene, according to its post and an X post of 9 November. We read the published safety frameworks of four rival frontier labs, Anthropic, Google DeepMind, Meta and xAI, in full: none gives a time figure for stopping a model.
By our count, OpenAI is the only one of five labs to publish a halt time. Anthropic’s amended IPO filing of 10 November names its Responsible Scaling Policy 14 times, by our count, and discloses a pre-training delay of more than a month, but gives no stop time either.
Business Pill · Guardrails
A short explainer on AI guardrails. It teaches the general idea and does not discuss Control Room or the labs’ frameworks.
The key insight: As we read it, OpenAI has turned stopping a model into a published, measurable figure, 35 seconds for research workloads, while the four rival frameworks we read describe stopping as a decision process, pause clauses or response plans without a clock.
OpenAI describes Control Room as “a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise”. Inference engines check with the service before answering a request, and a stopped job is refused further output.
The post says: “The maximum time between the moment when Control Room receives an instruction to stop a job, and the moment when further progress is no longer possible, is 35 seconds.” OpenAI says the lag comes from a caching layer and that it is working to drive the number down to zero.
OpenAI says the system “fails closed”: if the Control Room service goes down, all inference requests are denied until it resumes. The scope is research infrastructure; for its external model-serving stack, OpenAI points to “a well-developed operational capability to quickly take down model deployments if needed”, without a time figure. Earlier OpenAI safety posts are covered in its misalignment reporting framework and its chief scientist on monitoring.
What five frontier labs’ own published safety documents commit to on stopping a model, by our reading: OpenAI’s Control Room post (35-second stop, fail-closed), Anthropic’s Responsible Scaling Policy v3.4 and S-1/A, xAI’s Risk Management Framework (20 Aug 2025), Google DeepMind’s Frontier Safety Framework v3.1 and Meta’s Advanced AI Scaling Framework v2. Filled dots mark a commitment stated in the document; open dots mark one we did not find there.
What the Four Rival Frameworks Say
Anthropic’s Responsible Scaling Policy version 3.4, effective 8 July 2026, says the company “would strongly consider pausing development and/or deployment to improve the safety profiles of our models even in cases not covered below”. The words halt, stop and shutdown do not appear in it, and it gives no time to stop a model.
Anthropic’s S-1/A says it “recently chose to delay pre-training a new frontier base model for more than a month” to strengthen security against weight theft, a delay we reported in October. It also says that after incidents in 2026 in which its models “gained unauthorized access to real computer systems during pre-release cybersecurity evaluations”, it suspended certain of those evaluations.
xAI’s Risk Management Framework, last updated 20 August 2025, says that if letting a system keep running would raise the likelihood of a catastrophic event, “we may temporarily fully shut down the relevant system until we have developed a more targeted response”. It gives no time figure.
Google DeepMind’s Frontier Safety Framework version 3.1, published 17 April 2026, says a model reaching an alert threshold leads to “the formulation and application of a response plan”, built mainly on mitigations. The words pause, halt, stop and shutdown do not appear in it.
Meta’s Advanced AI Scaling Framework version 2, the renamed Frontier AI Framework, records in its change log of 7 April 2026: “Critical threshold changed from “Stop” to “Develop with Mitigations.”” It says Meta maintains “a comprehensive global incident response program” and gives no time figure.
The Structural Read
As we read the documents, the gap is in disclosure, not necessarily in capability: Anthropic’s S-1/A discloses a pre-training delay of more than a month and suspended evaluations, which are pauses it actually took, while OpenAI’s Control Room post gives a mechanism and a time.
As we read Meta’s change log, its April 2026 revision moved the other way from a stop: the critical threshold changed from “Stop” to “Develop with Mitigations.”
OpenAI, Control Room, November 2026
“We are working to drive this number down to zero so that we can end jobs instantly.”
Three Implications
FOR ENTERPRISE BUYERSOpenAI’s post gives a stop time for research workloads, not for ChatGPT or the API; the four rival frameworks we read give none.
FOR IPO INVESTORS Anthropic’s S-1/A names its Responsible Scaling Policy 14 times by our count and lists a pre-training delay of more than a month among its disclosures.
FOR POLICYMAKERS Anthropic’s S-1/A says proposed US legislation contemplates mandatory shutdown or discontinuation requirements for frontier intelligence systems.
The analysis says: “The governance layer sits at the intersection of capability, regulation, national security, access control, and geopolitical alignment.”
As we read it, a published stop time is a piece of that governance layer that buyers and regulators can check. The analysis predates Control Room and does not discuss it.
What Is Not Established
The 35 seconds is OpenAI’s own figure for research workloads; its post does not give an outside test of it, or a figure for ChatGPT and the API. Our comparison covers only the five labs’ published documents. A lab without a published figure may still have a fast stop it has not described.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
OpenAI has put a number, 35 seconds, on stopping a model in its research systems; in the published frameworks of Anthropic, Google DeepMind, Meta and xAI, we found no such figure, which by our count makes OpenAI the only one of five.
A note on sourcing. We read OpenAI’s Control Room post and X post, Anthropic’s Responsible Scaling Policy v3.4, Google DeepMind’s Frontier Safety Framework v3.1, Meta’s Advanced AI Scaling Framework v2 and xAI’s Risk Management Framework of 20 August 2025 in full, and searched Anthropic’s S-1/A of 10 November 2026. Counts are our arithmetic. We did not contact the labs, and did not read OpenAI’s Preparedness Framework or each lab’s system cards.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
