If you have ever talked to a voice assistant and felt like you were playing a game of "your turn, my turn," OpenAI just changed the rules. On July 8, 2026, OpenAI announced GPT-Live, a new voice model that powers ChatGPT Voice and can listen and speak at the same time. The model also delegates hard questions to GPT-5.5, OpenAI's most capable reasoning model, while keeping the conversation flowing. More than 150 million people already use ChatGPT Voice every week, and this update replaces the underlying engine for all of them.
The shift from turn-based to full-duplex voice is the biggest architectural change to conversational AI since chatbots went mainstream.
What is GPT-Live and what changed on July 8?
GPT-Live is a family of voice models that OpenAI launched on July 8, 2026. Two versions went live immediately: GPT-Live-1 for paying subscribers (Go, Plus, and Pro tiers) and GPT-Live-1 mini for free users. Both are now the default model powering ChatGPT Voice on iOS, Android, and the ChatGPT website.
To understand why this matters, you need to know how voice AI worked before. A "voice model" is an AI system that can understand spoken words and respond with spoken words. Previous voice assistants used one of two approaches.
A cascaded voice system (the original ChatGPT Voice) chained three separate models together. First, a speech-to-text model transcribed your words into text. Then, a large language model, or LLM, generated a text response. Finally, a text-to-speech model converted that text back into audio. Each step took time, and meaning could get lost between models. Think of it like a game of telephone: you speak, one system writes it down, another system thinks about it, and a third system reads the answer out loud.
A turn-based voice model (ChatGPT Advanced Voice Mode, launched in 2024) processed and generated audio inside a single model, which made responses faster and smoother. But it still waited for you to stop talking before it started responding. If you paused to think, the model might jump in and interrupt you. If there was background noise, it might mistake that noise for the end of your turn.
GPT-Live solves both problems with what OpenAI calls a full-duplex architecture.
How does full-duplex voice actually work?
"Full-duplex" is a term from telecommunications. A full-duplex phone line lets both people talk and listen at the same time, unlike a walkie-talkie where only one person can speak at a time. GPT-Live applies the same idea to AI: the model processes your incoming audio continuously while generating its own outgoing speech.
Instead of treating each conversation as a sequence of separate messages (your turn, then my turn), GPT-Live runs continuously. The model makes interaction decisions many times per second: whether to speak, keep listening, pause, interrupt you, or invoke a tool. This is why it can say "mhmm" or "got it" while you are still talking, letting you know it is following along without breaking your flow.
The chart below shows how the number of model components changed across the three voice architectures OpenAI has shipped.
| Architecture | How it works | Key limitation |
|---|---|---|
| Cascaded (original) | 3 models chained: speech-to-text, LLM, text-to-speech | Slow, lossy, stilted |
| Turn-based (Advanced Voice Mode) | 1 model handles audio, but waits for your turn to end | Rigid, interrupts on pauses |
| Full-duplex (GPT-Live) | 1 voice model runs continuously, can delegate to GPT-5.5 | No video or screen sharing yet |
The trade-off matters. A cascaded system used three models but produced laggy, robotic conversations. A turn-based system used one model but still felt like a walkie-talkie. GPT-Live uses one model for conversation and can call in a second model for heavy lifting, which brings us to the most important change.
What does delegating to GPT-5.5 mean in practice?
This is the feature that matters most for builders. When you ask GPT-Live a simple question like "what time is it," the voice model handles it directly and instantly. But when you ask something that requires web search, complex reasoning, or multi-step work, GPT-Live delegates the task to GPT-5.5 in the background.
"Delegation" means the voice model sends your question to a separate, more powerful model, gets the answer back, and weaves it into the conversation. While GPT-5.5 is working on the hard part, GPT-Live can keep talking with you about something else. When the answer is ready, GPT-Live brings it into the conversation naturally.
You can also choose the reasoning level. "Reasoning" in this context means how much time the model spends thinking before answering. Three levels are available to paying users:
- Instant: Uses GPT-5.5 Instant for fast responses. Good for quick questions.
- Medium: Uses GPT-5.5 Thinking with medium reasoning effort. The model spends more time on harder problems.
- High: Uses GPT-5.5 Thinking with high reasoning effort. Best for complex analysis or research.
Free users get GPT-Live-1 mini, which uses the Instant level only. OpenAI says they will swap in newer frontier models as they release them, so the background intelligence upgrades without you changing anything.
This architecture is what the AI industry calls "agentic": the voice model acts as a front-end orchestrator that can call other models and tools as needed. If you want to understand how this pattern compares to standard API integrations, our MCP explainer for software engineers breaks down what a protocol-based approach adds beyond a normal API call.
Did OpenAI prove GPT-Live is better than the old voice mode?
OpenAI ran human evaluations comparing GPT-Live to Advanced Voice Mode across five measures: overall preference, turn-taking, interruptions, conversational flow, and how natural the interaction felt. In matched 5 to 10 minute conversations, human testers strongly preferred GPT-Live-1 and GPT-Live-1 mini over Advanced Voice Mode.
On three benchmark tests, GPT-Live-1 also came out ahead:
- GPQA: Tests expert-level reasoning in biology, chemistry, and physics. GPT-Live-1 substantially outperformed Advanced Voice Mode.
- BrowseComp: Tests whether an AI agent can find hard-to-locate information on the web. GPT-Live-1 showed strong gains.
- tau3-Voice Telecom: Tests voice agents on realistic, multi-turn telecom support tasks. GPT-Live-1 outperformed Advanced Voice Mode.
OpenAI did not publish exact scores for these benchmarks, so we know GPT-Live won but not by how much. The human preference evaluations are the more meaningful result for everyday users: if people feel the conversation is more natural, that is what drives adoption.
What should you do with GPT-Live right now?
For most beginners, the answer is simple: open the ChatGPT app, tap the Voice button, and try it. The experience is live now for all users globally. You can test the difference between reasoning levels by asking the same complex question at Instant and then at High to see how the depth of the answer changes.
For builders and coders, the situation is more complicated. The GPT-Live API is not available yet. OpenAI says developers and enterprises can sign up to be notified, but there is no release date. If you are building a voice app today, you still need to use existing speech-to-text and text-to-speech APIs, or wait.
That said, the delegation architecture tells you where voice AI is heading. The pattern is: a fast, always-on model handles conversation, and a slower, smarter model handles intelligence. If you are designing a voice product, you should plan for this split. Your conversation layer needs to be fast and responsive. Your intelligence layer needs to be deep and accurate. Coupling them too tightly is the mistake the cascaded architecture made, and you should avoid repeating it.
A few caveats worth knowing. GPT-Live does not yet support voice with video or screen sharing in ChatGPT. Some languages may have a non-native accent or gaps in fluency. You can still access the older Advanced Voice Mode and Standard Voice if you need those features.
On safety, OpenAI says GPT-Live includes safeguards that can act while the model is speaking, including steering away from unsafe content or ending conversations in high-risk situations. The system card, available through OpenAI's announcement page, details testing across self-harm, emotional reliance, violence, and other risk areas. Parents can control whether teens use Voice through Parental Controls. Given that 150 million people use Voice weekly, the safety design is worth reading if you are building anything in this space.
The conversation layer and the intelligence layer are now separate
The architectural shift in GPT-Live matters more than its conversational smoothness. OpenAI split the conversation layer from the intelligence layer. The voice model handles the rhythm of human interaction. The frontier model handles the thinking. They talk to each other, but they do not need to be the same thing.
If you are building with AI, this is the pattern to watch. A single model that does everything is the old design. A fast model that orchestrates slower, smarter models is the new one.
Sources
- OpenAI - Introducing GPT-Live
