OpenAI just released GPT-Live, a new voice model designed to make conversations with AI feel genuinely human-like rather than robotic or scripted. The model can listen and speak simultaneously, interrupt users mid-sentence, and stay silent when appropriate, marking a significant shift in how voice assistants interact with people.

What Makes GPT-Live Different From Previous Voice Assistants?

For years, AI voice assistants have struggled with a fundamental problem: they sound artificial and interrupt constantly. GPT-Live addresses these issues through what OpenAI calls a "full-duplex architecture," which means the model can listen and speak at the same time. This is the technology that powers phone calls and real conversations, but it's rarely been implemented in AI assistants until now.

The model includes subtle human touches that make interactions feel more natural. During conversations, GPT-Live sprinkles in acknowledgments like "right" or "mmhmm" to show it's paying attention, and it can pause for longer periods when users need time to think. It also understands context well enough to know when to stop talking or let someone finish their thought.

"You can even forget you're talking to an AI. How should I put it? It's a really, really amazing feeling," said Yuchen Zhang, a research engineer at OpenAI, during a livestreamed demo on Wednesday.

Yuchen Zhang, Research Engineer at OpenAI

OpenAI demonstrated the model's capabilities by testing it in Apple CarPlay, where a user asked it to interrupt after the tenth letter while reciting the alphabet. The model succeeded, a task that would confuse most existing voice systems. This real-time responsiveness represents a major leap forward in voice AI usability.

How to Use GPT-Live Across Your Devices?

• Mobile Access: GPT-Live is rolling out to ChatGPT users on iOS and Android starting today, with the ability to set different intelligence levels for different conversations.
• Web Platform: The voice mode is also available on the web version of ChatGPT, making it accessible from any browser without needing to download an app.
• CarPlay Integration: ChatGPT's CarPlay support now includes the new voice mode, allowing drivers to have natural conversations while keeping their hands on the wheel.
• API Access: Developers interested in integrating GPT-Live into their own applications can express interest through OpenAI's API program, with full API support coming soon.

What Are the Two Versions of GPT-Live?

OpenAI is launching two versions of the model to serve different user needs. GPT-Live-1 is the full-featured version available to all paid ChatGPT subscribers, while GPT-Live-1-mini is a lighter version designed for free users. Both models are rolling out globally starting today.

For more complex reasoning tasks, GPT-Live-1 defers to GPT-5.5, OpenAI's more powerful reasoning model. This hybrid approach allows the voice model to handle casual conversation while still accessing deeper analytical capabilities when needed. The model also includes upgraded real-time web search and translation features, making it useful for tasks beyond simple chitchat.

Who Is OpenAI Targeting With This Technology?

OpenAI's marketing approach hints at a broader vision for voice AI. In promotional materials, the company featured elderly women using GPT-Live for practical tasks like getting knitting tips, checking public transit delays, and translating languages. This choice suggests OpenAI sees voice assistants as potential solutions to loneliness and isolation, not just convenient ways to get information.

"This model is one step closer to a truly accessible AGI, a world where talking to AI actually starts to feel like a real conversation," said Kundan Kumar, a research engineer at OpenAI.

Kundan Kumar, Research Engineer at OpenAI

OpenAI frames GPT-Live as progress toward artificial general intelligence, or AGI, a hypothetical AI system that could perform any intellectual task a human can do. The company's emphasis on natural conversation suggests it believes human-like interaction is a key component of truly accessible AI.

What Safety Measures Does OpenAI Include?

OpenAI has built safeguards into GPT-Live to prevent some of the harms caused by earlier voice assistants. The model will not imitate the voices of real people, addressing concerns that arose when an earlier OpenAI voice assistant sounded similar to actress Scarlett Johansson. If the model detects users discussing dangerous subjects like self-harm or violence, it will surface information about health and safety resources and may end the conversation.

The company notes that GPT-Live has been optimized for popular languages, though some languages may have non-native accents or fluency gaps. OpenAI says it is actively working to improve the experience across languages. One limitation at launch: the voice mode does not yet support video or screen sharing in ChatGPT, though these features are in development.

The rollout of GPT-Live represents a significant moment for voice-based AI interaction. By solving the technical challenges of natural conversation, OpenAI is betting that the future of AI assistants lies not in typing at a screen, but in talking to something that actually listens back.