Taking a New Model to Launch in ChatGPT: Readiness, Staged Rollout and Monitoring

You are a PM at OpenAI, and a new model is ready to ship in ChatGPT. How would you launch it? ```hint Plan for who could lose People build habits and workflows around a model's behavior. Think about which users could be worse off after the change, and what your launch plan owes them. ``` ### Clarifying Questions - Is the new model meant to replace the current default for everyone, or to sit alongside it for specific tasks such as reasoning or coding? - Which tiers get it first: Free, paid individual plans, or business and education workspaces with admin controls? Does the API launch at the same time? - How does its serving cost compare with the current default, and is there enough capacity to give it to everyone on day one? - Is a fixed date or a competitor's launch driving the timing? ### What a Strong Answer Covers - Launch goals and go/no-go criteria tied to what the model improves, agreed before launch. - Readiness work: capability and regression evaluations, safety testing, and capacity and cost planning. - A staged rollout across tiers, and a plan for the previous model and the users who depend on it. - Communication to consumers, business customers and developers, plus support and feedback channels. - Launch monitoring with guardrails, rollback triggers and post-launch follow-through. ### Follow-up Questions - In the first week, heavy users say the new model feels worse for their workflows even though evaluation scores went up. What do you do? - Launch-day demand exceeds compute capacity. How do you decide who gets the model and on what limits? - How would you measure the model's real effect when the launch is a public announcement that every user knows about?