OpenAI Launches Control Room: Research Workloads Halted Within 35 Seconds

What happened

Stop instructions now execute at the inference layer: Control Room connects research workloads, inference engines and a central service; each job registers and receives an identifier when it starts, and an engine must obtain authorization before returning model output; once a job is marked stopped, its next request is denied and the engine returns a refusal instead of model output.

Thirty-five seconds is the ceiling set by caching: an authorization cache supports the inference traffic volume, and the longest gap between a stop instruction and further progress becoming impossible is 35 seconds; OpenAI says it is working to reduce the delay to zero. The system works with misalignment monitoring, so critical alerts can automatically pause an affected workload, and a person can also trigger a stop.

Fail-closed already covers frontier workloads: fail-closed behaviour is live for all frontier-model research workloads, with the remaining research workloads to follow over the next few days; OpenAI also runs agentic and programmatic sweeps to catch engine implementations launched with Control Room missing or switched off.

Key facts

1. Announced by: OpenAI, Nov 9, 2026, on its official blog and X account
2. Stop ceiling: at most 35 seconds from a stop instruction to model output being halted, with the delay coming from the authorization cache
3. Enforcement point: authorization at the inference layer rather than researcher code stopping itself; each job must be tied to an approved safety case
4. Fail-closed: if the service goes down or is tampered with, all inference requests are denied; live for frontier research workloads, with the rest to follow within days
5. Scope: all inference workloads across the research infrastructure, including agentic sampling for training and evaluation and data processing; the external model-serving stack is outside this deployment
6. Background: in a Sep 20 internal research-agent incident, monitors flagged the problem within 15 minutes, but a person stopped the run manually two and a half hours later