OpenAI launches GPT-Live full-duplex voice for ChatGPT

Every prior version of ChatGPT Voice was a walkie-talkie: you spoke, it waited for you to stop, then it answered. GPT-Live breaks that turn-based contract. It is a full-duplex model, meaning it can speak and listen at the same time ^[1]. Instead of one big decision per turn, it makes interaction decisions many times per second - whether to speak, keep listening, pause, interrupt, or invoke a tool ^[2]. That is why it can drop in short backchannel cues like 'mhmm' or 'yeah' while you are still talking, and why it can absorb an interruption without restarting its whole answer ^[2]. The constant timing awareness is also what unlocks live simultaneous translation: in press briefings the model spoke a running translation as the presenter talked, something the old turn-based system structurally could not do ^[3].

The design decision underneath the demo is delegation. GPT-Live is not trying to be the smartest model in the room; it is trying to be the fastest conversational surface. When a task needs web search, deeper reasoning, or agentic work, GPT-Live hands it off to a frontier model such as GPT-5.5 running in the background while it keeps the conversation flowing ^[2]. In OpenAI's flagship demo the voice model chatted naturally while concurrently correcting a historical date, checking transit delays, and pulling weather. This is a two-tier architecture - a low-latency talker in front, a heavyweight reasoner behind - and it is the same split OpenAI is drawing across its product line, where GPT-Live plays for voice the front-end role that its coding surfaces play for agentic work.

The launch messaging is unambiguous. Sam Altman said the model feels 'magical and real' and that he expects his own preference to shift from typing to talking, and the official rollout leans on that framing. In matched test conversations, GPT-Live-1 and its mini variant were strongly preferred over Advanced Voice Mode ^[2], and TechRadar billed it as ChatGPT's smartest voice model ever ^[6]. But the same feature that makes it feel alive is also the source of the loudest real-world gripe.

Across Reddit's OpenAI, ChatGPT, and singularity communities, reception has been positive-with-caveats: users call it the closest thing yet to the movie 'Her' and praise how it listens while speaking and runs tool calls in the background, but the single most-repeated complaint is that the model interrupts too eagerly, sprinkling filler backchannels like 'Awesome,' 'Mhmm,' and 'Wow' into the conversation. A contrarian minority argues it is not a real leap and still carries the same uncanny quality, and independent early-access reviewers were more skeptical than the polished official videos, confirming the duplex overlap works only 'sometimes.' The tension is instructive: backchanneling is a feature engineered to make the model sound human, and it is also the behavior users most want to turn down. That is not a bug report so much as a calibration problem baked into the full-duplex bet.

GPT-Live arrived globally on iOS, Android, and web, but developers cannot build on it yet: there is no day-one API, with access planned later via a signup form ^[2]. That is a deliberate ordering. OpenAI is putting voice in front of 150 million-plus people who already use ChatGPT's Voice and Dictation features ^[1]before it opens the model to third parties, and it frames voice as a potential primary interface to computing for managing increasingly complex long-running agentic work ^[1]. The consumer-first sequencing signals that OpenAI sees the strategic prize as the everyday interaction layer, not the API line item.

The cost of that ordering is competitive exposure. Google's Gemini Live already offers full-duplex conversation plus camera and screen sharing, capabilities GPT-Live lacks at launch ^[5], and a field of specialized vendors - ElevenLabs, Deepgram, Cartesia, InWorld AI - already ship developer-facing, low-latency voice products ^[5]. Because enterprises cannot yet wire GPT-Live into their own voice agents, that penetration is expected to lag rivals with existing developer products ^[5]. Latency is the battleground: OpenAI's rollout reportedly targets roughly 300ms for real-time voice ^[7], while competitors advertise aggressive numbers, with InWorld AI claiming sub-250ms P90 and Cartesia publishing time-to-first-byte around 40ms for a turbo voice model ^[5]. The translation demo also showed the polish gap: TechCrunch observed the Hindi output carried a heavy American accent and a slightly bookish tone ^[1].

GPT-Live is the end of a multi-year compression story, and reading the arc explains why full-duplex feels inevitable in hindsight. The original ChatGPT Voice was a cascade of three separate systems: Whisper transcribed your speech to text, GPT-4 reasoned over the text, and a text-to-speech model spoke the answer back ^[1]. Each hop added latency and stripped away nuance, because the reasoning model never actually heard your voice. In May 2024, GPT-4o collapsed that stack into a single model trained end-to-end across text, vision, and audio ^[2], and the Advanced Voice Mode rollout that followed put one natively-audio model in users' hands, first on mobile in mid-2024 and then on the web later that year ^[3].

Even that unified model, though, still operated in discrete turns - it had to wait for you to stop speaking, which produced stiff back-and-forth and mistimed interruptions ^[5]. The 2025 gpt-realtime release pushed speech-to-speech into production voice agents via the Realtime API ^[4], and GPT-Live is the next step: it removes the turn boundary itself. So the trajectory reads cleanly - three models to one model, then one model that spoke in turns to one model that listens and speaks continuously. Notably, the reasoning did not necessarily get smarter at the voice layer; skeptics point out core reasoning is unchanged and the model can still miscount letters. What changed is the interface, which is exactly what OpenAI is betting matters most.