Summary
OpenAI’s DevDay 2026 session introduces GPT Live 1, a full duplex voice model that streams audio in and out at the same time, so it can listen and speak simultaneously. Presenters Alex, an engineer on GPT Live, and Erika, who works with startups building voice applications, are joined onstage by a voice agent co-host named Marin, who controls the demos and diagrams entirely by voice.
The session walks through three demos that illustrate the core concepts for building with the model. The first covers model personality, using a second GPT Live session with a grumpy British voice named Colin to play a game of Fortunately, Unfortunately. It shows how the model handles turn-taking, back-channel reactions, and interruptions on its own, without a separate voice activity detector. The presenters also note that simulated second sessions can serve as test users for evaluating voice apps.
The second demo covers delegation, in which GPT Live keeps talking to the user while a more powerful backend model or agent harness does complex work. A live drawing-guessing game uses a split architecture of UI, voice interface, and a backend running code execution. The talk explains how to feed background state to the voice model through silent context updates and speakable commentary. The final demo looks at the turnless architecture and how to apply it in your own applications.
📺 Source: OpenAI · Published November 11, 2026
🏷️ Format: Keynote Launch
