Product Launch · XiaoHu Explainer

OpenAI Launches GPT-Live: No More Taking Turns—Hard Questions Go Straight to GPT-5.5

A true duplex voice model is live today; complex tasks can be handed off to GPT-5.5 in real time.
30-Second Overview
  • OpenAI has released its next-generation voice model, GPT-Live, built on a full-duplex architecture that keeps processing input while generating speech — enabling natural conversations where it listens and speaks at the same time.
  • When a question needs web search, deep reasoning, or a complex task, GPT-Live hands it off in real time to GPT-5.5 running in the background, so your conversation never stalls.
  • Two versions — GPT-Live-1 and GPT-Live-1 mini — roll out today to ChatGPT users worldwide: Go / Plus / Pro default to GPT-Live-1, while Free users default to the mini version.
  • In human head-to-head evaluations, both GPT-Live-1 and mini were clearly preferred over Advanced Voice Mode, and both also outperform it on GPQA, BrowseComp, and τ³-Voice Telecom.
  • Voice mode now shows visual cards for weather / stocks / sports, lets you adjust reasoning effort (Instant / Medium / High), and adds strengthened safety protections around self-harm and emotional dependence tailored to voice.
This article is an explainer based on OpenAI's official launch blog post. The evaluation results, preference data, and safety-testing conclusions cited here all come from OpenAI's own self-reported testing and have not yet been independently verified by a third party.
1Launch

OpenAI Launches GPT-Live: Voice Conversations No Longer Have to Take Turns

On July 8, 2026, OpenAI released its next-generation voice model, GPT-Live, positioning it as the new underlying model for ChatGPT Voice. It's rolling out to all users worldwide starting today.

What GPT-Live does is let you talk to AI without taking turns: it can listen while you're speaking and respond at the same time — you can interrupt it mid-sentence or jump in with a follow-up. When it hits a question that needs a web search or some real thinking, it quietly hands the work off to the more powerful GPT-5.5 running in the background, and your side of the conversation never breaks stride.

Why it matters: The previous Advanced Voice Mode had to wait until you finished talking before it could respond. GPT-Live achieves listen-while-speaking through a full-duplex architecture, and it beats Advanced Voice Mode on all three of GPQA, BrowseComp, and τ³-Voice Telecom. Both GPT-Live-1 and GPT-Live-1 mini roll out globally starting today.

OpenAI's official launch demo video (with Chinese/English subtitles): a genuine voice conversation between GPT-Live and a "grandma" persona, used to showcase the full-duplex traits of listening while speaking, interrupting at will, and waiting quietly. Source: OpenAI, "Introducing GPT-Live."
Your voiceAI's voice

Two waveforms can rise and fall at the same time without interrupting each other — that's what full-duplex conversation looks like.

2The Old Approach

Where the Old Approach Fell Short: Relay Handoffs Lose Information, Turn-Taking Causes Interruptions

To see what's actually new about GPT-Live, it helps to know where the previous two generations of voice tech got stuck. Both could hold a conversation with you, but each had its own awkward flaw. Laying all three architectures side by side makes this clearest.

Gen 1
Three models in relay
(cascaded system)
You speak Speech to text AI thinks of an answer Text to speech
Each step has to wait for the last one to finish, and tone or intonation can get lost at each handoff — replies come out slow and stiff.
Gen 2
One model sends and receives,
but still takes turns
One model handles
both directions
Waits for silence
= assumes you're done
Then it replies
Faster and smoother, but only one side can talk at a time; if you pause to think, or there's background noise, it can easily mistake that for "you're done" and cut in.
GPT-Live
Listen while speaking
(full-duplex)
Keeps listening Keeps responding Decides multiple times a second
Listening and speaking happen at once, constantly judging whether to talk, listen, wait, or let itself be interrupted — closing the gaps left by the first two generations. The next two sections dig into how it pulls this off.
An Analogy · Cascaded System

It's like a relay race: speech-to-text, answer generation, and text-to-speech are the three runners, and every handoff risks dropping the baton — meaning a bit of what you said gets lost along the way.

Here's what the same line — "hey, got a minute to chat" — sounds like across all three generations. OpenAI provided real recordings you can compare directly:

Gen 1 · Cascaded System
Slow and stiff responses, with long pauses. (Standard Voice, using GPT-5.5 Instant)
Gen 2 · Turn-Based (AVM)
Faster and smoother, but multi-turn exchanges still feel stiff. (ChatGPT Advanced Voice Mode)
GPT-Live · Full-Duplex
Fast, natural, expressive responses, and it listens more attentively too. (GPT-Live-1, using GPT-5.5 Instant)
3Breakthrough One · Full-Duplex

Breakthrough One: One Model Listens and Speaks at Once, No Waiting for You to Finish

GPT-Live's first change is a full-duplex architecture (meaning it can listen and speak at the same time) built specifically for ongoing conversation. Instead of breaking a conversation into separate messages, it continuously processes your input while continuously generating a response.

Core Innovation

Because listening and speaking run simultaneously, the model can make many decisions every second: speak now, keep listening, pause for a beat, interrupt, or call a tool. The result is a more natural back-and-forth with a better sense of timing — it can even do real-time translation.

OpenAI's official demo video (with Chinese/English subtitles): an OpenAI employee explains the "continuous presence" capability and demonstrates it live by having GPT-Live call out the time throughout a pour-over coffee routine — in a multi-step task, the model stays with you the whole way and can jump in anytime. Source: OpenAI, "Introducing GPT-Live."
An Analogy · Full-Duplex

A walkie-talkie only lets one person talk at a time while the other waits; a phone call lets both sides jump in and pick up on each other anytime. A full-duplex voice model is AI switching from walkie-talkie mode to phone-call mode.

Walkie-Talkie Mode · Old Version

You talk → release → then it talks → it finishes → then it's your turn again. Pause in the middle to think, and it's easily mistaken for "done talking," so it jumps in and interrupts.

Phone-Call Mode · GPT-Live

It's listening and judging the whole time you're talking. When you pause to think something through, it waits quietly; when you signal it to jump in, it does — and it'll drop in an "mm-hm" or "got it" so you know it's still following along.

OpenAI's official demo video (with Chinese/English subtitles): a real conversation sample showing GPT-Live-1 (backed by GPT-5.5 Instant) listening and speaking at once — a direct contrast to the walkie-talkie mode shown above. Source: OpenAI, "Introducing GPT-Live."
Under the full-duplex architecture, these are the calls the model is repeatedly making every second
Speak Keep listening Pause briefly Interrupt Call a tool
4Breakthrough Two · Delegation

Breakthrough Two: It Handles Everyday Chat Itself, Hands Off Hard Problems to a Stronger Model

The second change is splitting "chatting with you" apart from "doing the heavy lifting." GPT-Live focuses on keeping the conversation going; when a question needs a web search, real reasoning, or a chain of agentic actions, it delegates that hard task to a more powerful model — at launch, that's GPT-5.5This backend engine isn't hardwired: whenever OpenAI ships a stronger frontier model down the line, whatever sits behind GPT-Live simply gets swapped out — no rework needed at the interaction layer.. While the delegation is happening, your side of the conversation keeps going as normal, and the result gets folded back in once it's ready.

Core Innovation

Here's a concrete scenario showing how this plays out: while driving, you ask ChatGPT by voice, "Can you check if there are still seats on tomorrow's Beijing-to-Shanghai flights?" That needs a web lookup, so GPT-Live keeps chatting with you while handing the task to the background; once the background finishes, the answer gets folded into the conversation, and you never feel any lag.

Foreground · GPT-Live (conversation never stops) Background · GPT-5.5 (search · reasoning) You ask "Are there seats on tomorrow's flight?" GPT-Live keeps chatting "Let me check — what time are you leaving?" GPT-Live reports back "3 flights still have seats, earliest one's at 8am" conversation continues GPT-5.5 searches flight availability online delegate task bring back result
The conversation never stalls the whole time: during the few seconds the background is checking flights, GPT-Live up front isn't stuck — it's still confirming your departure time with you. Once the result's in, it flows in naturally and gets spoken. This split also means GPT-Live's backend engine can keep upgrading right along with future models.
Real Conversation Recording · Delegating a Deep Task
GPT-Live delivers fast, natural responses while GPT-5.5 handles the search task in the background. (GPT-Live-1, using GPT-5.5 Instant) Source: OpenAI, "Introducing GPT-Live."
5Evaluations

Evaluation Results: Better on Both Chat Feel and Hard Questions Than the Old Version

For this launch, OpenAI ran new human evaluations to measure how comfortable and smooth conversations feel, and also benchmarked against Advanced Voice Mode on several general-capability evals. All the results below come from OpenAI's own reporting.

How the human head-to-head evaluation worked: people had 5-to-10-minute conversations and compared them across five dimensions — overall preference, turn-taking, handling interruptions, fluency, and naturalness — rating each conversation as preferring either GPT-Live or Advanced Voice Mode.

Overall preferenceTurn-takingHandling interruptionsFluencyNaturalness

Head-to-head preference rate (50% means a tie with AVM):

GPT-Live-1
75.7%
GPT-Live-1 mini
69.2%

There's also a separate rated conversation-quality test (out of 7 points, covering conversational flow and overall comfort):

ModelConversational FlowComfort
GPT-Live-14.965.19
GPT-Live-1 mini4.334.47
Advanced Voice Mode3.803.82

These three general-purpose evals weren't designed specifically for voice — they test the model's underlying capability — and were also benchmarked against Advanced Voice Mode:

GPQA
Tests expert-level scientific reasoning across biology, chemistry, and physics.
BrowseComp
Tests the ability to hunt down hard-to-find information on the web (agentic web search).
τ³-Voice Telecom
An in-house OpenAI eval that has a voice agent handle realistic, multi-turn phone customer-service tasks.

GPQA accuracy, broken down by the reasoning effort of the backend model:

AVM
45.3%
GPT-Live-1 mini
74.9%
GPT-Live-1 · Instant
76.5%
GPT-Live-1 · Medium
81.7%
GPT-Live-1 · High
84.2%

BrowseComp accuracy:

AVM
0.7%
GPT-Live-1 mini
31.6%
GPT-Live-1 · Instant
35.1%
GPT-Live-1 · Medium
60.6%
GPT-Live-1 · High
75.2%

The τ³-Voice Telecom eval looks at both task completion rate and how long tasks take — higher reasoning effort is more accurate but slower:

ModelMedian Task DurationTask Success Rate
AVM385.5 sec29.5%
GPT-Live-1 mini290.9 sec39.5%
GPT-Live-1 · Instant231 sec37.3%
GPT-Live-1 · Medium295 sec59.6%
GPT-Live-1 · High386 sec63.4%
Which engine's behind it: GPT-Live-1 (instant) and mini both run on background GPT-5.5 Instant; GPT-Live-1 Medium and High run on GPT-5.5 Thinking, at medium and high reasoning effort respectively. The τ³ eval was run using a custom user model powered by OpenAI's latest reasoning model; all data comes from OpenAI's own reporting.
6In Practice

Open ChatGPT Voice Now, and Here's What You'll Notice

More than 150 million people use ChatGPT's voice and dictation features every week — as a hands-free daily assistant, to practice a foreign language, to tell bedtime stories, or just to chat during a commute. Starting today, tapping the voice button means you're talking to GPT-Live.

OpenAI's official demo video (with Chinese/English subtitles): several OpenAI employees discuss the model, cutting to two real-world scenes — hands-free voice editing of an offer-letter reply, and using voice to look things up mid-dance without touching the phone. Source: OpenAI, "Introducing GPT-Live."
OpenAI's official demo video (with Chinese/English subtitles): the actual startup experience after tapping the ChatGPT voice button. Source: OpenAI, "Introducing GPT-Live."
150M+ /week
People using ChatGPT voice and dictation every week
2 versions
GPT-Live-1 (default for Go / Plus / Pro) and mini (default for Free)
9 voices
All voice options rebuilt from scratch for this release
GPT-5.5
The frontier model handling deep tasks in the background (Instant / Thinking)

More Like Talking to a Real Person, and a Better Listener Too

You can jump in with a question anytime, pause to gather your thoughts, or ask it to slow down. It'll respond with things like "mm-hm" or "got it" to show it's following along. All nine voices have also been rebuilt from scratch.

Old Version · When You Pause

If you're just pausing to think about wording, it easily assumes you're done and jumps straight in, cutting you off.

GPT-Live · When You Pause

It waits quietly instead of rushing to reply. Tell it to stay quiet and just listen, and it will. With background noise like traffic or people talking nearby, it's also better at staying focused on your voice.

Smarter Answers When You Need Them

Voice mode can now call on the latest frontier model, and you can pick the reasoning effort yourself: use Instant when you want speed, or Medium / High when you want it to think things through more.

Instant
Fastest answers, good for everyday quick questions.
Medium
Spends a bit more time thinking, balancing speed and depth.
High
Use when you want it to really think it through — longest thinking time.
OpenAI's official demo video: the actual interface for choosing reasoning effort (Instant / Medium / High) during a voice conversation. Source: OpenAI, "Introducing GPT-Live."

Some Answers Are More Useful When You Can See Them

As it talks, ChatGPT can now show you cards directly — weather, stocks, sports scores, and the like — without you having to type to confirm. Voice mode still supports search, memory, and image and file uploads as usual.

OpenAI's official demo video (with Chinese/English subtitles): while packing a suitcase, someone asks "can you show me the weather in Mexico City this week?" and ChatGPT pops up a weather card directly, no need to stop and type to confirm. Source: OpenAI, "Introducing GPT-Live."

Here are two more real screenshots from OpenAI:

ChatGPT Voice showing a card with upcoming international soccer match schedules
Example of a sports schedule card
ChatGPT Voice showing a map card of nearby bars
Example of a map card (nearby bars to watch the game)
7Safety

Safety Design Built Specifically for Voice

On top of the safety foundation of the latest models, GPT-Live adds dedicated safety training and protections tailored to this new voice format.

Safety Testing Closer to Real-World Use

OpenAI extended its safety testing to audio-native evaluations, built a batch of tests using synthetic audio to specifically target several key risk categories, and had internal experts run red-team testing focused on voice-specific risks. According to OpenAI, GPT-Live matched or outperformed Advanced Voice Mode across nearly every evaluation area. The key risk categories covered include:

Self-harmPsychosis and maniaEmotional reliance on AIViolenceSexual content

Protections That Can Step In Mid-Speech

Because voice conversation happens in real time, OpenAI built protections that can kick in while the model is still speaking. Once the system detects potentially unsafe content, it can take one of three actions:

Action 1
Steer the model toward a safer response
Action 2
Add extra safety guidance or support resources
Action 3
For higher-risk cases, end the voice conversation directly

For conversations involving self-harm, OpenAI has adapted ChatGPT's support flow for the voice format, including expert-reviewed crisis hotline support.

Teen and Voice Protections

For teen users, age-appropriate behavior was built into the model's training. Parents can use Parental Controls to decide whether their teen can use ChatGPT Voice at all; in high-risk situations showing signs of self-harm or suicidal intent, the linked parent account may be notified. GPT-Live also only uses a set of preset voices with built-in safeguards, and won't mimic a real person's voice.

8Rollout

Can You Use It Now? Availability and Limitations

GPT-Live is rolling out to ChatGPT users worldwide starting now, covering iOS, Android, and ChatGPT.com. Which version each plan defaults to is laid out in the tier breakdown below.

GPT-Live-1Default for Go / Plus / Pro
The standard version, now the default voice model for Go, Plus, and Pro users.
GPT-Live-1 miniDefault for Free
The lightweight version, now the default voice model for Free users.
Legacy Standard / Advanced Voice Mode
Still available as a manual option. Features like voice paired with video or screen sharing, which GPT-Live doesn't yet support, are still available there.

A few current limitations worth knowing: GPT-Live is optimized for ChatGPT's most widely used languages, and some languages may still come with a non-native accent or feel less fluent — OpenAI says it's working on this. At launch, it doesn't yet support voice paired with video or screen sharing, though those are coming soon. An API is also on the way — developers and businesses can sign up via a form to get notified.

  • For everyday voice Q&A, language practice, or chatting during a commute, interruptions, thinking pauses, and requests to slow down are all recognized naturally — no more getting talked over or misread as often.
  • For questions needing a web search, expert-level science knowledge, or complex multi-step tasks, voice conversations can now get near-deep-model accuracy, since GPT-5.5 is handling it in real time behind the scenes.
  • Checking weather, stocks, or sports scores now shows a visual card directly in the conversation, no need to type to confirm.
  • Parents can use Parental Controls to manage whether a teen can use voice features, and high-risk conversations showing signs of self-harm or suicidal intent will notify the linked parent.
GPT-Live is built on a full-duplex architecture, meaning it can listen and speak at the same time. For questions that need web search, deeper reasoning, or more complex work, it delegates to our latest frontier model behind the scenes, then brings the result back into the conversation once it's ready. OpenAI, "Introducing GPT-Live," July 8, 2026
This article is an explainer based on OpenAI's official launch blog post, "Introducing GPT-Live" (July 8, 2026). The evaluation data, preference results, and safety-testing conclusions cited here all come from OpenAI's own official reporting and have not yet been independently verified by a third party. Original post: openai.com/index/introducing-gpt-live/