In current AI voice interactions, traditional cascading architectures often cause noticeable delays due to the sequential processing of speech-to-text, model inference, and text-to-speech. To completely address this pain point, OpenAI has officially launched a new full-duplex speech model, GPT-Live-1, in its API, bringing a human-like smooth voice interaction experience directly to developers.

Traditional voice agents often struggle with complex conversation scenarios, especially when dealing with pauses, interruptions, or topic changes. GPT-Live-1, however, uses a single model to process both input and output audio, achieving a revolutionary simplification in architecture. It not only supports immediate response interruptions but also seamlessly delegates deep reasoning and tool calls to backend text models, allowing conversations to continue in the background while maintaining high flexibility.

In terms of actual performance and application scenarios, this model brings several core advantages. First, it has excellent interruption handling capabilities. Early tests showed that the language learning platform Speak found it reduced interruptions during thinking pauses by nearly 80%. Second, developers can customize the agent's tone, rhythm, and conversational style through system prompts, and it can perfectly handle background noise and silent contexts, making it ideal for telephone customer service, restaurant reservations, and long conversation interactions. In benchmark tests, it performed exceptionally well when paired with GPT-6 Astra, which has moderate reasoning intensity.

Currently, GPT-Live-1 is fully available in the API, with a front-end speech layer pricing of $0.05 per minute. Several industry partners, such as Yelp and Intercom, have stated that this model significantly improves the accuracy of turn switching, allowing AI voice support to truly move beyond the stop-and-start rhythm, achieving a new experience of natural and smooth conversation similar to face-to-face interactions with humans.