Skip to main content
EN · FR Start for free
Menu
Lesson 6 of 12 · ChatGPT in 2026: From GPT-5.6 to GPT-6

GPT-Live: The Voice That Listens and Speaks Simultaneously

Lesson 6 of 12 Updated November 5, 2026 Intermediate
On this page
  1. Three Generations of Voice
  2. Intelligence in the Background
  3. Availability at Launch
  4. What This Changes for Your Use Cases
  5. Your Turn
  6. Key Takeaways

Status as of November 5, 2026. GPT-Live 1 has been available in the API since September 10, 2026, at $0.05 per minute of voice session. Two French voices, Brise and Sillage, were added on October 16 (Lesson 10). In ChatGPT, the Live option now runs on GPT-Live-1.1 and GPT-Live-1.1 Quip. Since September 9, the voice mode can delegate difficult questions to GPT-5.6 or GPT-6 Astra.

On July 8, 2026, OpenAI launched GPT-Live, a new generation of voice models that now powers ChatGPT Voice. This isn’t just an upgrade to the existing engine; it’s a fundamental architectural shift. To understand what this new system brings, we need to look at the architectures that came before it.

Three Generations of Voice

The first version of ChatGPT Voice chained three models together: one to transcribe your speech into text, a large language model to generate the response, and a text-to-speech engine to synthesize the audio. This cascade let you talk to an advanced model for the first time, but it lost information at every step, making responses slow and choppy.

Advanced Voice Mode later processed and generated audio within a single model, reducing latency. However, it still operated turn-by-turn: the model waited for you to finish speaking before it responded. Since the end of a turn was detected by silence, a short pause or background noise could trigger the model to interrupt at the wrong moment.

GPT-Live uses a full-duplex architecture: it continuously processes what it hears while it speaks, deciding multiple times per second whether to talk, listen, pause, interrupt itself, or call a tool. In practice, it can show it’s following along (“mhmm,” “yes”) without cutting you off, handle rapid exchanges, or stay silent while you take time to think.

Three Generations of Voice Three columns. Cascade: transcription, language model, then text-to-speech; slow responses and information loss. Turn-by-turn: single audio model, smoother, but waits for the turn to end and misinterprets pauses. Full-duplex, GPT-Live: listens and speaks simultaneously, delegating difficult work to a more powerful model. FROM CASCADE TO FULL-DUPLEX Three Generations of Voice Cascade Principle Three models in a chain Limitation Slow, losses between links Turn-by-turn Principle Single audio model Limitation Waits for turn to end Full-duplex Principle Listens and speaks at the same time Advantage Delegates difficult work
Based on the announcement "Introducing GPT-Live" (OpenAI, July 8, 2026).

Intelligence in the Background

GPT-Live doesn’t handle difficult tasks alone. When a question requires web search, deep reasoning, or a complex task, it delegates the work to a more powerful model and brings the result back into the conversation once ready. Meanwhile, it keeps talking to you; the conversation doesn’t freeze. At launch, this background model was GPT-5.5, and OpenAI stated it would adopt new models as they released. Since September 9, ChatGPT Voice can rely on GPT-5.6 or GPT-6 Astra.

Availability at Launch

ElementAs of July 8, 2026
ModelsGPT-Live-1 for paid plans, GPT-Live-1 mini for the Free plan
ChatGPTDeployed on chatgpt.com and iOS/Android apps in supported regions
Business, Enterprise, and Edu WorkspacesNot available at launch
Video and Screen SharingNot supported; remained available via Advanced Voice Mode
APIAnnounced for later, with a sign-up form

On July 31, OpenAI added a SynthID watermark to supported audio content generated with GPT-Live, both in ChatGPT and the API, along with a provenance verification tool.

What This Changes for Your Use Cases

For training, voice-based customer support, or hands-free assistance, full-duplex removes the main pain point of voice assistants: rigid turn-taking. A trainer using voice to simulate an interview can now recover the rhythm of a real conversation, complete with interruptions, follow-ups, and silences. Today, you can find these settings under Settings > Voice, which offers three options: Live, Advanced Mode, and Standard Mode. Video and screen sharing on mobile still require Advanced Mode.

Your Turn

Test continuous conversation. In ChatGPT Voice with the Live option enabled, try these three scenarios: interrupt the response mid-sentence, take a long pause to think, and ask a question that requires a web search while continuing to talk as it executes. Note how the conversation reacts each time, then repeat the test using Advanced Mode to compare.

Official Sources: Introducing GPT-Live (OpenAI, July 8, 2026) · ChatGPT release notes · ChatGPT Voice (OpenAI Help Center) · GPT-Live 1 (OpenAI API Documentation)

Key Takeaways

  • GPT-Live is full-duplex: it listens and speaks simultaneously without waiting for your turn to end
  • It delegates difficult questions to a more powerful model and continues speaking in the meantime
  • At launch: GPT-Live-1 for paid plans, GPT-Live-1 mini for Free, with no video or professional workspaces
  • The API opened on September 10, 2026; in ChatGPT, configure Live under Settings > Voice
← Previous lessonProgrammatic Tool Calling: Model-Orchestrated ToolsNext lesson →ChatGPT in July 2026: Work, Plugins, Copilot, Presence, and Health