Introducing GPT-Live

OpenAI dropped GPT-Live this week, and the name alone tells you where the company's head is at. This isn't another chat model with a text-to-speech bolt-on. It's positioned as a dedicated generation of voice models built specifically for the back-and-forth of real conversation, and it's already live inside ChatGPT Voice.

For years, voice assistants have felt like talking to a phone tree with better manners. You pause, it processes, it responds a beat too late or cuts you off mid-thought. OpenAI's pitch here is that GPT-Live closes that gap, handling the timing, tone, and flow of human speech in a way that earlier voice stacks never quite managed. That's the whole game in voice AI: not what the model says, but when and how it says it.

OpenAI has been circling this problem since the GPT-4o voice demos in 2024, when live demonstrations showed a model that could laugh, interrupt itself, and shift tone mid-sentence. Those demos were flashy but the shipped product lagged behind the hype for months. GPT-Live reads like the company finally closing that distance between demo reel and daily-use product, baking the responsiveness into the core model rather than stitching it together after the fact.

What's notable is the framing as infrastructure, not a feature. OpenAI isn't calling this a ChatGPT update. It's calling it a new class of model, which suggests they expect GPT-Live or its descendants to show up well beyond the ChatGPT app — in customer service lines, cars, earbuds, wherever a company decides typing is the wrong interface. And given how much OpenAI has leaned into voice as a wedge for daily habitual use, that ambition tracks.