Voice model that listens and talks at once hits the API

On September 10, OpenAI opened GPT-Live-1 to API developers. This is the full-duplex voice model that has run ChatGPT Voice since July 8. It listens while it speaks, so a caller can cut into an agent mid-sentence and the conversation keeps going.

According to the OpenAI announcement, the API release adds stronger instruction following, custom voices and telephony support. OpenAI says interruptions are handled better because a single model reasons over incoming and outgoing audio together. When a request needs reasoning or tool calls, the model can hand that work off to backend systems.

Full duplex means the model takes in audio and produces audio continuously instead of taking turns. When the model launched in ChatGPT in July, Tech Times reported that OpenAI described it deciding “many times per second” whether to speak, keep listening, pause, acknowledge with a brief “mhmm,” or invoke a background tool. GPT-Live isn’t built to do complex reasoning itself. In ChatGPT, web search, deeper reasoning and multi-step analysis go to GPT-5.5, which runs in the background while the conversation continues. Because of that split, the backend handoff in the API release matters more than its one line in the announcement suggests.

The API release is the latest step in a thread that has been reworking turn-taking since 2023:

• 2023: The original ChatGPT voice chains three models: Whisper for speech-to-text, GPT-4 for the response, and a text-to-speech engine. Each stage waits for the one before it, and tone, pacing and intonation are lost along the way.
• 2024: Advanced Voice Mode moves to a single audio-native model that keeps prosody (the rhythm and intonation of speech) but still works in turns. It treats silence as the end of a turn, so a pause or background noise can cause an unwanted interruption or an awkward gap.
• July 8: GPT-Live replaces Advanced Voice Mode in ChatGPT. GPT-Live-1 becomes the default for Go, Plus and Pro subscribers, and GPT-Live-1 mini becomes the default for free users. Tech Times put the switch at more than 150 million people who use ChatGPT’s voice and dictation features every week.
• July 31: SynthID watermarking is added to GPT-Live audio.
• September 10: GPT-Live-1 reaches the API with custom voices, telephony support and handoff of reasoning and tool calls to backend systems.