OpenAI recently announced that it has officially made the GPT-Live-1, which is currently used for the voice function of ChatGPT, available to developers. Developers can now integrate this voice model into their own applications and business processes through an API, building full-duplex voice assistants that can listen to users in real time and respond simultaneously. The front-end voice processing cost of this model is $0.05 per minute, while the back-end model used for complex reasoning will be billed separately according to the corresponding model price.

image.png

Listen and speak at the same time, naturally handling interruptions and pauses

GPT-Live-1 is OpenAI's new generation real-time voice model launched this year, with the main feature being a full-duplex architecture. Traditional voice AI usually completes three steps: speech recognition, language model reasoning, and speech synthesis, and the system can only process and generate answers after the user finishes speaking. However, GPT-Live-1 can generate voice output while listening to the user's speech, so it can more naturally handle interruptions, pauses, agreement, and fast back-and-forth conversations, as well as determine when to continue speaking, pause, listen, or call tools.

OpenAI stated that this architecture significantly reduces the delay in voice interaction and solves the connection problems that often occur in traditional pipelines. Developers no longer need to coordinate complex switches between multiple independent models themselves. In testing, GPT-Live-1 outperformed GPT-Realtime-2.1 by 30 percentage points on the Full Duplex Bench, especially showing significant improvements in turn-taking delay and interactive behavior. When paired with GPT-6 Astra at medium reasoning intensity, it also ranked first in the Tau3 test, which evaluates end-to-end voice agent tasks.

Complex reasoning is handled by the back-end model, with costs that can be combined as needed

GPT-Live-1 does not handle all complex reasoning. OpenAI separates the voice model responsible for real-time listening and speaking from the back-end model responsible for deep reasoning: when users ask questions that require searching, complex analysis, or longer tasks, GPT-Live-1 will hand over the work to the back-end model while continuing to maintain natural voice communication. Therefore, developers can choose back-end models based on their business needs—using faster and cheaper models for simple tasks, and stronger reasoning models for complex issues.

In terms of application, GPT-Live-1 supports phone scenarios natively, and can be used for full-duplex voice agents such as restaurant reservations and customer support. In early cases, the language learning platform Speak found that compared to previous round-based systems, GPT-Live-1 reduced the system's active interruption of learners by nearly 80%. The model also provides native voice recognition transcription and response text, supporting keyword bias and turn detection, and is optimized for noisy environments. OpenAI also expanded the available voices, covering a wider range of accents, dialects, and languages. Developers can also combine GPT-Live-1 with tools like Codex, where the voice model receives tasks, the back-end tools process them, and then return to the voice conversation.