OpenAI’s GPT-Live lets ChatGPT listen and speak at once

TL;DR:

  • OpenAI has launched GPT-Live, a new generation of voice models now powering ChatGPT Voice.
  • A full-duplex architecture lets it listen and speak at the same time, making decisions many times per second rather than waiting for the user to stop.
  • For hard questions it delegates to a frontier model in the background while keeping the conversation flowing.

OpenAI has launched GPT-Live, a family of voice models built to make talking to AI feel like a real conversation rather than a series of stilted exchanges. Rolling out to ChatGPT users globally from today in two variants, GPT-Live-1 and a smaller “mini” version, the models replace the older approach that powered ChatGPT’s voice mode and are described by OpenAI as its smartest voice system yet.

Full-duplex, with a frontier model behind it

The core change is architectural. Earlier voice systems either chained separate speech-to-text, language and text-to-speech models together, losing information and adding lag, or processed audio in a single model but still waited for discrete turns, so a brief pause or background noise could trigger an awkward interruption. GPT-Live instead uses a full-duplex design that processes input continuously while generating output, deciding many times a second whether to speak, listen, pause or invoke a tool. It can offer conversational cues like “mhmm”, handle live translation, or simply stay quiet while a user thinks. A second change decouples that natural interaction from deeper work: when a query needs search or reasoning, GPT-Live delegates to a frontier model, GPT-5.5 at launch, and folds the answer back in without breaking the flow.

Looking forward

For UK businesses, the significance is less the novelty than the trajectory. OpenAI frames the release as a step towards voice becoming a practical interface for “longer-running and more agentic work”, a shift from voice-as-gimmick to voice-as-workflow in call centres, accessibility tools and hands-free professional settings. It also intensifies a fast-moving frontier week: GPT-Live lands the same day as SpaceXAI’s Grok 4.5 and just ahead of OpenAI’s own delayed GPT-5.6 public launch. API access is promised soon, at which point the real test, how naturally these agents handle interruption, latency and background reasoning in live deployments, begins.

Share this article