GPT-Live makes ChatGPT voice a continuous conversation

OpenAI introduced GPT-Live-1 and GPT-Live-1 mini on July 8 and began a global rollout in ChatGPT. The models support full-duplex interaction, which allows the system to keep listening while it speaks and respond to interruptions without forcing every exchange into a strict call-and-response pattern.

OpenAI said the voice system can pass search, reasoning and other complex work to GPT-5.5 while maintaining the conversation. That division lets the live model handle timing and speech while another model works on a heavier task. How well the handoff preserves context, discloses delays and avoids speaking over a user remains an operational question rather than a settled launch claim.

API access was planned for later and was not presented as generally available with the first consumer release. Developers should therefore separate the ChatGPT rollout from any future programmable access. Plan eligibility, supported languages and regional availability can also change and need confirmation in current release notes.

A July 31 update added SynthID support for eligible generated audio, along with public and API verification routes. The addition came after the initial launch and should be described as a later provenance update, not a feature necessarily present in every audio output from day one.

The key test is not whether a demonstration sounds human. A continuously listening assistant needs to recover from interruptions, remain quiet at the right moment, manage background speech and translate without losing meaning. TechCrunch noted unnatural Hindi in a demonstration, but one example is not a multilingual evaluation and cannot support a broad quality ranking.