OpenAI's Control Room blocks model outputs from stopped research workloads within 35 seconds, even if its own service fails.