OpenAI introduces GPT-Live

Until now, voice conversations with AI have largely followed a turn-based pattern: the user speaks, the AI listens, and only then responds. While effective, this approach has always felt less natural than a real human conversation.

With the introduction of GPT-Live, OpenAI is changing that. The company has launched a new voice model that powers ChatGPT Voice, enabling the AI to listen, respond, and handle complex tasks simultaneously. Instead of waiting for users to finish speaking, GPT-Live can participate in conversations much more naturally.

The biggest innovation behind GPT-Live is its full-duplex architecture. Unlike previous voice models that alternated between listening and speaking, GPT-Live can do both at the same time, making conversations feel significantly more fluid. For example, the model can:

avoid interrupting users during pauses;
acknowledge what it hears with brief responses such as “mm-hmm” or “I understand”;
continue following the conversation even if the user changes direction mid-sentence;
perform real-time language translation.

According to OpenAI, the model evaluates the conversation multiple times per second, continuously deciding whether it should listen, speak, wait, or avoid interrupting the user.

GPT-Live is not responsible only for maintaining conversations. When users ask questions that require web browsing, deep reasoning, or complex analysis, GPT-Live automatically delegates those requests to OpenAI’s most capable reasoning models. Currently, these background tasks are handled by GPT-5.5.

The key difference is that users don’t have to wait in silence. GPT-Live continues the conversation naturally while the background model processes the request, then seamlessly incorporates the final answer into the discussion once it’s ready.

GPT-Live also brings significant improvements to ChatGPT Voice. The updated experience includes:

more natural and expressive conversations;
the ability for users to interrupt the AI at any time;
better handling of long pauses without ending the conversation;
improved performance in noisy environments;
visual cards for weather, sports scores, stock prices, and other real-time information.

Users can also choose between different response modes, including Instant for faster replies and Medium or High for more thoughtful reasoning.

OpenAI says GPT-Live includes a new set of safety mechanisms designed specifically for real-time voice interactions. The model can:

detect conversations involving self-harm;
modify potentially harmful responses in real time;
direct users to appropriate support resources when necessary;
end voice conversations in certain high-risk situations.

According to OpenAI, GPT-Live has also been evaluated across a wide range of challenging scenarios, including psychological risks, violent content, and sexually explicit material. GPT-Live is rolling out gradually to users worldwide.