OpenAI's new voice model lets apps listen and talk at the same time

Voice bots have felt broken for years because they are built like a relay race. OpenAI's new GPT-Live-1, just opened to outside developers, replaces the three-model handoff with a single network that listens and speaks in the same pass. The relay is the standard recipe today: a speech-to-text model transcribes the caller, a text-based reasoning model decides what to say, and a text-to-speech model reads the answer out loud. Each handoff adds delay, and each handoff is a place where timing, tone, or context can break.

GPT-Live-1 was first built into ChatGPT, and the API release gives any developer the same surface, with a reference page covering capabilities and parameters. For a builder, the practical change is the end of the relay. Instead of stitching three models together and tuning each handoff, they write a system prompt that sets the agent's tone and pace, and they get a voice surface that can hold a turn, pause naturally, and recover when the caller interrupts. For callers, the change is the disappearance of two familiar glitches: the half-second pause before a bot "realizes" you spoke over it, and the brittle turn-taking that forces you to wait for a full stop before the bot will respond. The model also handles background noise and silence without narrating every step, so a caller on a windy street does not hear the bot say "one moment" every few seconds.

When an agent needs to look something up, run a calculation, or reason through a multi-step problem, GPT-Live-1 can delegate that to a backend text model running in parallel and speak a short, natural response in the meantime. The voice layer stays thin, the reasoning layer stays swappable, and the caller does not hear the seam. OpenAI's post names "GPT-6 Astra" as one example of a future backend target, though that model is not shipping today.

The one quantified win in the launch is from Speak, a language-learning app that OpenAI cites as an early evaluator. Speak says interruptions in its Live Tutor lessons fell by almost 80% after switching to a full-duplex setup, an attribution Unite.ai's re-report traces to Speak co-founder Andrew Hsu. That figure is worth holding carefully: it is one partner's own measurement, on one product, using OpenAI's framing, and Speak has not published a methodology or a counter-benchmark. Treat it as evidence the architecture is moving in the right direction, not as a category-wide proof point.

This is not a consumer phone line. OpenAI's post frames full-duplex telephony as a developer capability, the kind of plumbing a reservation system, a clinic's after-hours line, or a customer-support vendor would build on top of the API. Codex and ChatGPT already ship with GPT-Live-1, which is some signal that the model holds up at product scale, even if it is not a published benchmark.

Pricing, per a third-party re-report, lands at roughly $0.05 per minute of audio. That number is not in OpenAI's announcement or on its published pricing page, so a builder planning a budget should confirm it against OpenAI's own rate card before committing. At a few cents a minute, a ten-minute support call costs under a dollar of inference, which is the kind of number that makes always-on voice agents worth a product team's time.

Google's Gemini Live and a handful of startups are also shipping full-duplex, single-model voice, so GPT-Live-1 is arriving into a category that has already started to converge on the same architecture. The next eighteen months of voice-AI releases will all be answering the same question this launch is meant to answer: can a model carry a real conversation, or is it still taking careful turns?

The first third-party benchmark that runs GPT-Live-1 against a chained pipeline, in public, will be the one worth watching.