|
ResearchAudio.io
GPT-Live Talks While GPT-5.5 Thinks
Full-duplex voice that listens and speaks at once. The two-model split, explained.
|
|
150M+
talk to ChatGPT by
voice each week
|
9
voices, remastered
for GPT-Live
|
2
models shipping:
GPT-Live-1 and mini
|
|
|
For three years, talking to an AI meant taking turns. You spoke, it waited for silence, then it replied. OpenAI's new GPT-Live drops the turn: it listens and speaks at the same time, and when a question needs real thinking, it hands that off to GPT-5.5 in the background while it keeps the conversation going.
|
Why turns were the bottleneck
|
|
The first version of ChatGPT Voice, launched in 2023, chained three separate models together. A speech-to-text model transcribed what you said, a language model wrote a reply, and a text-to-speech model read it back. It was the first time anyone could hold a spoken conversation with a frontier model, but the handoffs were costly. Detail got lost between models, and replies came back slow and stilted.
|
|
The generation that followed, Advanced Voice Mode, folded all of that into a single model. Latency dropped and the voice sounded more human, but the interaction still ran on discrete turns. The model had to wait for you to stop talking before it could respond, and it decided you were done by listening for silence. A pause to gather your thoughts, or a bit of street noise, could be read as the end of your turn, so the model would cut in at the wrong moment.
|
|
GPT-Live changes the shape of the problem with two architectural moves.
|
|
Three eras of voice AI
|
2023 · Cascaded
Three models in a chain
stt →
llm →
tts
Slow, stilted. Detail lost between models.
|
→ |
2024 · Turn-based
One model, one turn
listen …wait…
speak
Smoother, but waits for silence, then cuts in on a pause.
|
→ |
2026 · GPT-Live
One continuous stream
listen ⇄ speak
Full-duplex. Natural, real-time, delegates to GPT-5.5.
|
Source: OpenAI, "Introducing GPT-Live" (July 8, 2026)
|
One: it never stops listening
|
|
The first change is a full-duplex architecture. Instead of processing a sequence of separate messages, GPT-Live processes incoming audio and generates outgoing audio at the same time. That lets it make an interaction decision many times per second: keep listening, start speaking, pause, cut in, or reach for a tool.
|
|
Think of the difference between a walkie-talkie and a phone call. Turn-based voice is a walkie-talkie, where one side transmits while the other waits. Full-duplex is a phone call, where both people can talk, pause, and react in real time. In practice this is what lets GPT-Live drop in small acknowledgments like "mhmm" or "yeah," hold a quick back-and-forth, or stay quiet when you clearly need a second to think. OpenAI says the same capability gives the model a better sense of timing and lets it do live translation.
|
Two: it delegates the hard part
|
|
The second change is a split. OpenAI separated the model that runs the live conversation from the model that does the deep work. When you ask something that needs a web search, careful reasoning, or a multi-step task, GPT-Live delegates that job to a stronger model, GPT-5.5, and keeps talking to you while the answer is computed. When the result lands, it folds back into the conversation.
|
|
This decoupling has a second payoff. Because the conversational layer sits apart from the reasoning layer, OpenAI can swap in newer frontier models behind GPT-Live over time without rebuilding the voice experience. At launch it runs on GPT-5.5 in the background. GPT-Live-1 (instant) and GPT-Live-1 mini use the GPT-5.5 Instant model, while the Medium and High settings use GPT-5.5 Thinking with more reasoning effort.
|
|
How delegation keeps the conversation going
|
Foreground · never stops
|
You
|
⇄ |
GPT-Live · listens + speaks
"mhmm", quick back-and-forth
|
|
|
hard query handed down ▼ ▲ result returns
|
|
Background · heavy lifting
|
GPT-5.5
search + reason
|
→ |
Answer folds back in
spoken mid-conversation
|
|
|
What the evaluations show
|
|
OpenAI built new human evaluations to measure how pleasant a conversation felt and how well it flowed. In matched five to ten minute conversations, testers preferred GPT-Live-1 and GPT-Live-1 mini over Advanced Voice Mode on overall preference, turn-taking, interruptions, flow, and how natural each exchange felt.
|
|
On capability tests, OpenAI reports that GPT-Live-1 substantially outperforms Advanced Voice Mode on GPQA (expert-level science across biology, chemistry, and physics), shows strong gains on BrowseComp (agentic web search for hard-to-find information), and comes out ahead on an internal telecom-support benchmark called τ³-Voice Telecom. One caveat worth flagging: OpenAI described these results in relative terms and did not publish specific scores in the announcement, so treat the size of each gain as directional rather than measured.
|
|
The key idea: GPT-Live separates interaction from intelligence. The voice model's job is to be a good conversational partner in real time. The frontier model's job is to be smart. Keep the two apart, and the voice layer inherits every future reasoning upgrade automatically.
|
|
|
|
|
GPT-Live is rolling out to ChatGPT users worldwide across iOS, Android, and the ChatGPT web app. GPT-Live-1 becomes the default model behind ChatGPT Voice for Go, Plus, and Pro users, and GPT-Live-1 mini becomes the default on the no-cost tier. You can pick a reasoning level to match the moment: Instant for fast replies, Medium or High when you want more thinking time.
|
|
Voice also gains visual cards. While you talk, ChatGPT can show a weather forecast, a stock chart, sports fixtures, or a map, and it still supports search, memory, images, and file uploads. Two limits at launch: GPT-Live does not yet work with video or screen sharing, and some languages may have a non-native accent or gaps in fluency. The legacy Standard and Advanced Voice Modes stay available where those features are needed.
|
|
Safety built for real-time voice
|
|
|
Because voice unfolds live, some safeguards run while the model is speaking. If the system detects a risky response, it can steer the model toward a safer one, surface supportive resources, or end the conversation in higher-risk cases, with support flows adapted for spoken interaction and crisis helpline support. OpenAI added protections for teen users and trained age-appropriate behavior into the model, with parental controls over voice access. It also says GPT-Live is built for conversation, not impersonation: a fixed set of voices, plus safeguards against imitating a real person. OpenAI reports the model performed comparably to or better than Advanced Voice Mode across nearly all of the areas it evaluated, with details in the GPT-Live system card.
|
|
|
|
The headline is that GPT-Live feels natural to talk to, but the durable move is the split between conversation and reasoning. Voice is the interface that most exposes latency, because people notice a delayed reply far more in speech than in text. By letting a fast, always-listening model hold the floor while a slower model thinks, OpenAI keeps the conversation fluid without capping how hard the system can reason. It also turns the background model into a swappable part, which is where the compounding happens.
|
|
|
|
The bet is that delegation stays invisible. If GPT-5.5 takes a few seconds on a hard query, can a stream of small acknowledgments really hide the gap, or will people learn to hear the seams. And once GPT-Live reaches the API, what does a voice agent that can pause, cut in, and think in the background make possible that turn-based agents could not. That is the part worth watching.
|
|
"Voice AI stopped taking turns. GPT-Live listens and speaks at once, and delegates the hard thinking to GPT-5.5 in the background."
|
|
|
Next issue: what full-duplex means for real-time voice agents once GPT-Live reaches the API.
ResearchAudio.io
Source: Introducing GPT-Live, OpenAI, July 8, 2026.
|
