How fast can OpenAI's Control Room stop a model?

OpenAI says at most 35 seconds pass between Control Room receiving a stop instruction and further model progress becoming impossible. The delay comes from a caching layer for inference authorization that handles the traffic volume. OpenAI says it aims to bring this lag down to zero so that jobs can be ended instantly.

What happens if OpenAI's Control Room service goes down?

OpenAI's Control Room fails closed. If the service is unavailable because of a system error, because it was disabled, or because someone tampered with it, every research inference request is denied until it recovers. OpenAI says this is already rolled out to all frontier model research workloads and will reach the remaining research workloads over the next few days.

Does OpenAI's Control Room apply to ChatGPT or the API?

No. OpenAI describes Control Room as a system for its research infrastructure: reinforcement learning experiments, evaluations, and large sampling or data processing jobs. Outside API traffic from Astra-class models is covered by OpenAI's separate misalignment monitoring for the Responses API, which Control Room works alongside.

Why did OpenAI put the stop switch in the inference engine?

OpenAI says workload code changes often and is spread across many teams, so it is a poor place for critical safety controls. There are only a small number of inference code paths, so enforcing Control Room there gives strong guarantees that every request is checked. Agentic and programmatic sweeps also look for engines launched without Control Room.

OpenAI · 2026-11-09 · major

OpenAI Control Room — a fail-closed off switch for research model workloads

In short

OpenAI Control Room tracks every research inference job and can stop one within 35 seconds. Inference engines ask it before each request, and if the service goes down, all research inference is denied.

OpenAI · 2026-11-09 · 2 sources checked

OpenAI artwork with the words Control Room among routed lines and pastel diamonds

OpenAI moved its emergency stop into the inference layer, so a halted research job cannot get another token from any model.

Quick facts

Maker	OpenAI
Published	9 November 2026
Scope	All research inference: RL sampling, evals, data processing
Stop time	Up to 35 seconds (target: zero)
Failure mode	Fails closed: no service, no inference
Rollout	All frontier research workloads now; the rest over the next few days

What is it?

Control Room is OpenAI's central service for tracking model workloads across its research infrastructure and stopping them when a safety or security concern comes up. It covers every research inference workload, including agentic sampling for training and evaluation, and it works alongside OpenAI's misalignment monitors. Every research workload that runs agentic inference must also be tied to an approved safety case.

How does it work?

A workload asks Control Room for a job_id when it starts and passes that id with every sampling request. The inference engine checks with Control Room before serving each request and refuses requests from jobs marked as terminated. When a training monitor spots a problem it sends a stop message with the job_id, and the job's next request to any model on any engine is denied. Regular sweeps look for engines running without Control Room.

Why does it matter?

Enforcement sits in the few inference code paths instead of fast-changing workload code, which OpenAI says gives strong guarantees that every request is checked. The design answers the kind of failure seen in OpenAI's September DNS incident, which a monitor caught. For other labs and companies running agents, it is a concrete pattern: one authorization service per job, checked at the model, that denies everything when it is unavailable.

Who is it for?

AI safety, platform and security engineers

Frequently asked questions

How fast can OpenAI's Control Room stop a model?: OpenAI says at most 35 seconds pass between Control Room receiving a stop instruction and further model progress becoming impossible. The delay comes from a caching layer for inference authorization that handles the traffic volume. OpenAI says it aims to bring this lag down to zero so that jobs can be ended instantly.
What happens if OpenAI's Control Room service goes down?: OpenAI's Control Room fails closed. If the service is unavailable because of a system error, because it was disabled, or because someone tampered with it, every research inference request is denied until it recovers. OpenAI says this is already rolled out to all frontier model research workloads and will reach the remaining research workloads over the next few days.
Does OpenAI's Control Room apply to ChatGPT or the API?: No. OpenAI describes Control Room as a system for its research infrastructure: reinforcement learning experiments, evaluations, and large sampling or data processing jobs. Outside API traffic from Astra-class models is covered by OpenAI's separate misalignment monitoring for the Responses API, which Control Room works alongside.
Why did OpenAI put the stop switch in the inference engine?: OpenAI says workload code changes often and is spread across many teams, so it is a poor place for critical safety controls. There are only a small number of inference code paths, so enforcing Control Room there gives strong guarantees that every request is checked. Agentic and programmatic sweeps also look for engines launched without Control Room.

Try it

Read the post: openai.com/index/introducing-control-room