Short Answer: What is GPT-Live?
GPT-Live is OpenAI’s full-duplex voice model — it listens and speaks at the same time instead of waiting for a user to finish talking. It has powered ChatGPT Voice since July 2026, and on September 10, 2026, OpenAI brought it to the API so developers can build their own voice agents on it.
Quick Summary
- GPT-Live launched July 8, 2026, replacing the old speech-to-text → LLM → text-to-speech pipeline behind ChatGPT Voice with a single full-duplex model.
- On September 10, 2026, OpenAI released GPT-Live-1 in the API at $0.05 per minute, letting any developer build on the same voice model.
- OpenAI never published an end-to-end latency number for GPT-Live. An independent lab test found it is actually slower to stop talking when genuinely interrupted — but far better at not stopping for noise that isn’t meant for it.
- Early adopters include Yelp (phone reservations), Speak (language tutoring), Cognition, Picsart, and Heygen.
- GPT-Live delegates harder reasoning to a backend model — GPT-5.5 at launch, now GPT-6 Astra — while keeping the conversation going.
OpenAI announced on September 10, 2026 that GPT-Live-1, the voice model behind ChatGPT Voice, is now available in its API. That makes this week’s actual news the developer release of a model that already reshaped ChatGPT’s voice mode two months earlier — not a brand-new voice AI. Here is what GPT-Live is, what changed this week, and what an independent lab found when it tested the marketing claims against a real phone.
What Is GPT-Live?
OpenAI introduced GPT-Live on July 8, 2026, as a replacement for the cascaded voice pipeline that had powered ChatGPT Voice since launch. The old system chained three separate models together for every exchange: a speech-to-text model to transcribe you, a language model to write a reply, and a text-to-speech model to read it back.
GPT-Live collapses that into one full-duplex model that listens and speaks at the same time, the way a phone call works. It can backchannel with cues like “mhmm” or “yeah,” go quiet while you’re thinking, and respond to being interrupted without a separate turn-detection system bolted on.
Two versions rolled out to ChatGPT globally: GPT-Live-1 as the default for paid tiers, and GPT-Live-1 mini for Free users, replacing Advanced Voice Mode. For anything requiring deeper reasoning or web search, GPT-Live hands the question to a frontier text model running in the background — GPT-5.5 at launch, since upgraded to GPT-6 Astra — and keeps talking while it waits for an answer.
What Changed on September 10?
GPT-Live-1 is now available in the OpenAI API, priced at $0.05 per minute for the voice layer. That means any developer, not just ChatGPT, can now build a voice agent on the same full-duplex model, and pair it with whichever backend model handles the reasoning and tool calls.
OpenAI highlighted a handful of capabilities specific to the API release:
- Interruption handling that reasons over incoming and outgoing audio together, instead of chaining separate speech-to-text, language, and text-to-speech models.
- Reasoning delegation to a backend model such as GPT-6 Astra or a third-party model for anything that needs real thinking.
- Tone and pace control through the system prompt, so developers can shape how an agent sounds.
- Telephony support, so a full-duplex voice agent can take real phone calls.
OpenAI cited Speak, a language-learning app, which found GPT-Live-1 cut conversational interruptions by almost 80% for learners compared with the previous turn-based system — giving students more room to think mid-sentence without the AI jumping in.
Does GPT-Live Actually Reduce Latency?
OpenAI’s own launch posts do not include an end-to-end latency figure for GPT-Live. Agora Media Lab ran its own controlled test in July, using real devices, an artificial mouth for repeatable speech input, and controlled network impairment — and the results complicate the “faster and more natural” framing.
When a user deliberately interrupts, GPT-Live was 500 milliseconds slower to go silent than the older, strictly turn-based Advanced Voice Mode. It spends that extra time deciding whether the interruption is real.
That trade-off buys something real, though. When Agora played sounds that were not meant as interruptions:
- GPT-Live correctly ignored all 30 background-speech probes.
- The older Advanced Voice Mode stopped talking for 20 of 30.
- The original cascaded Standard Voice Mode stopped for all 30.
In other words, GPT-Live is slower to yield to a genuine interruption but dramatically less likely to be derailed by a cough or someone else talking in the room. Agora also found a new failure mode the older systems never had: in a noisy room where nobody was addressing ChatGPT, GPT-Live spontaneously answered background voices in 4 of 30 ten-second test windows — a case of correctly hearing speech but misjudging who it was for.
The takeaway for anyone evaluating GPT-Live: the real improvement is not raw speed, it is judgment about when to talk. For more on what “inference” and model latency actually mean under the hood, see Prompt Insider’s explainer on what inference in AI is.
Who Is Already Building on GPT-Live-1?
OpenAI named several early API adopters at launch:
- Yelp is using it for Yelp Host, handling restaurant reservations by phone, including background noise and side conversations.
- Speak, a language-learning app, reported an 80% drop in mid-sentence interruptions for its AI tutor.
- Cognition, Picsart, and Heygen are also listed as early users, though OpenAI has not detailed their specific use cases.
One unnamed enterprise customer told OpenAI that switching from a cascaded voice build to GPT-Live-1 cut their codebase by 80% and removed 23,000 lines of code previously needed to coordinate the old three-model pipeline.
Why Does This Matter Beyond OpenAI?
Full-duplex, single-model voice architecture is a real shift from the cascaded pipelines every voice assistant has used for years, and OpenAI is not the only lab racing to ship it. The pattern echoes something Prompt Insider keeps running into with hardware and model launches generally: the marketing headline and the independently measured reality are two different stories. Apple’s A20 Pro doubled two different pieces of AI silicon this month, not one number restated twice; here, GPT-Live is not simply “faster,” it makes a different trade-off between speed and judgment. Both are worth understanding on their own terms before repeating the headline version.
Frequently Asked Questions
Is GPT-Live a new model released this week?
No. GPT-Live launched in ChatGPT Voice on July 8, 2026. What launched on September 10, 2026 is GPT-Live-1 in the OpenAI API, making the same voice model available to outside developers for the first time.
What does “full-duplex” mean for a voice AI?
It means the model can listen and speak at the same time, like a real phone call, instead of strictly waiting for one side to finish before the other responds. Older voice assistants are turn-based: they wait, then answer.
Is GPT-Live actually faster than the old ChatGPT Voice?
Not in every respect. Independent lab testing found GPT-Live is about 500 milliseconds slower to stop talking when a user deliberately interrupts it, compared to the previous Advanced Voice Mode. Its real advantage is filtering out sounds that are not genuine interruptions, which the older systems handled poorly.
How much does GPT-Live-1 cost in the API?
OpenAI priced the voice layer at $0.05 per minute for developers building on GPT-Live-1 through the API.
What model handles GPT-Live’s reasoning?
GPT-Live delegates harder questions, web search, and complex reasoning to a backend frontier text model while keeping the conversation going. At launch this was GPT-5.5; OpenAI has since updated it to GPT-6 Astra, and says it will continue updating the backend model over time.
Sources: OpenAI, OpenAI, Agora Media Lab.
About the author
Kai Williams
Kai Williams has been in marketing for years, with a long background in SEO before AEO had a name. He stepped into Answer Engine Optimization the moment AI started reshaping how people search, and has been tracking the shift ever since. At Prompt Insider, he covers AEO, AI marketing, and the future of search, breaking down what is changing and what brands need to do about it.
