Voice model that listens and talks at once hits the API

OpenAI opened GPT-Live-1 to developers through its API on September 10, 2026, a voice model built for full-duplex conversation that keeps listening while it speaks, with custom voices, telephony support and stronger instruction following added in the same release, according to the announcement published by OpenAI.

At a glance

• A single model reasons over incoming and outgoing audio at the same time, and according to OpenAI that is what improves interruption handling when a caller cuts in mid-sentence.
• GPT-Live first reached users in ChatGPT Voice on July 8, 2026, and OpenAI added SynthID watermarking to the model's audio on July 31, ahead of the developer release.
• GPT-Live-1 can also hand reasoning and tool calls off to backend systems, according to OpenAI, which describes that capability alongside the full-duplex audio in the API announcement.

Voice agents have mostly been assembled from a pipeline of speech recognition, a text model and a synthesizer, and each hop adds latency that callers hear as awkward pauses. Folding the audio into one model that runs in both directions reads as an attempt to remove that seam. The telephony hook suggests the target is call centre and support work rather than consumer demos.

Full duplex means the model does not wait for a turn to end. Because a single model reasons over the incoming stream and its own outgoing speech together, according to OpenAI, it can react to a caller who cuts in while it is still talking. OpenAI also states that the model can delegate reasoning and tool calls to backend systems.

Telephony support puts the model on phone infrastructure directly, and custom voices let a deployment use a voice of its own rather than a stock set. OpenAI names stronger instruction following alongside those two as part of what the API version brings.

GPT-Live reached consumers before developers. The model appeared in ChatGPT Voice on July 8, 2026, and SynthID watermarking was added to its audio on July 31, so the API version arrives about two months after the first public deployment of the model.

SynthID and the API audio

OpenAI's summary of the release does not name pricing, does not identify the telephony providers involved and does not say which regions are covered at launch. It also leaves open whether the SynthID watermarking applied to GPT-Live audio in ChatGPT travels with audio generated through the API. How access to custom voices is granted is not described either.