Part of our AI models explained: ChatGPT, Claude, Gemini and the rest guide
OpenAI says its new Control Room system halts a research model's output within 35 seconds of a decision to stop or pause a job. It is already deployed across OpenAI's research infrastructure and is built to fail closed: if Control Room goes down, inference requests are denied until it recovers.
How it works
In its Control Room post, published November 9, 2026, OpenAI describes three parts: the workload (a research job such as a reinforcement learning experiment, an evaluation or a large sampling run), the inference engine that runs the model on GPUs, and the central Control Room service.
- A job asks Control Room for a job ID. OpenAI says every research workload that does agentic inference must be tied to an approved safety case.
- Each time the job wants model output, it sends the job ID to an inference engine.
- Before answering, the engine asks Control Room whether that job is still allowed to run.
- If a person or a monitoring system has marked the job stopped, Control Room denies the request and the engine returns a refusal. The research process can stay alive, but it can no longer sample tokens.
OpenAI says all of its frontier reinforcement learning jobs run with a monitor: another model that inspects what the model being trained is doing and can raise an alert. It cites a recent DNS-related incident that a monitor caught.
What the 35 seconds means
| Property | What OpenAI says |
|---|---|
| Maximum delay | 35 seconds between Control Room receiving a stop instruction and further progress becoming impossible |
| Why it isn't instant | A caching layer for inference authorization, needed for the request volume; OpenAI says it is working to bring the delay to zero |
| If Control Room fails | Fail-closed: all inference requests are denied until it resumes |
| Rollout | Fail-closed is live for all frontier model research workloads and will reach the rest over the next few days |
| Scope | Research infrastructure; external model serving is described as more centralized, with an existing ability to take deployments down quickly |
The design choice is where the switch sits. OpenAI says workload code changes often and isn't a good place for critical protections, so enforcement is in the inference layer, where only a small number of code paths exist across its research. It also says it runs continual agentic and programmatic sweeps to check that nobody launches rogue engine implementations with Control Room disabled.
What it doesn't tell us
- No outcomes. The post gives no figures for how often jobs have been stopped, how many alerts were false alarms, or what Control Room has caught beyond the one incident it cites.
- Self-reported. This is OpenAI's description of its own system; no outside audit is cited in the post.
- Research, not customers. Control Room covers research workloads. The post does not say it governs products customers use.
Why it matters
Labs say their models are getting more capable, so how quickly a misbehaving job can be stopped is becoming a question in itself. The same logic applies to anyone running agents: Codema's guides to running coding agents safely with permissions and sandboxes and what Claude Code's auto mode approves cover the smaller-scale version of the same idea, with limits enforced outside the agent's own code. OpenAI's BioBasicsBench results are a reminder that model judgment still has gaps, and our explainer on agent harnesses covers why the layer around the model matters.
What to watch next
- The delay. OpenAI says it wants to push the 35-second lag down to zero.
- Rollout completion. Fail-closed was still reaching the remaining research workloads.
- Whether others publish similar controls, and whether any details of stop events are shared.
FAQ
What is OpenAI's Control Room?
It is a central service that checks whether a research job is still allowed to run. Before an inference engine answers a request, it asks Control Room, and if a person or monitor has marked the job stopped, the engine returns a refusal. The research process can stay alive but can no longer sample tokens.
Why does it take up to 35 seconds instead of stopping instantly?
OpenAI says the delay comes from a caching layer for inference authorization, which is needed for the request volume. It says it is working to bring the delay to zero.
What happens if Control Room goes down?
It fails closed: all inference requests are denied until it resumes. OpenAI says fail-closed is live for all frontier model research workloads and will reach the rest over the next few days.
Does Control Room govern products that customers use?
The post does not say so. Control Room covers research workloads, and OpenAI describes external model serving as more centralized, with an existing ability to take deployments down quickly.
Has Control Room been independently verified?
No outside audit is cited. This is OpenAI's description of its own system, and the post gives no figures for how often jobs have been stopped or how many alerts were false alarms. It cites one recent DNS-related incident that a monitor caught.
Sources
Topics
John F.AI desk
John F. is a Codema.ai house byline: a pen name used by our editorial desk, not one individual journalist. It covers the AI beat: new models and how they're priced, open models, AI coding tools and agents, and security incidents that affect AI developers. Pieces under this name are written with AI assistance and reviewed by a human editor before they're published.
All stories by John F.
