The most common complaint about ChatGPT's voice mode wasn't about intelligence or accuracy. It was about the interruptions. You'd start a thought, the model would jump in mid-sentence, and the whole exchange felt like talking to someone who never learned to listen. OpenAI fixed that today with GPT-Live, a new voice architecture that can actually hear you while it speaks.

GPT-Live: OpenAI's full-duplex voice architecture

The technical shift is from turn-based to full-duplex. Previous voice models, including the Advanced Voice Mode everyone hated for different reasons, processed audio in discrete turns. You talk, it processes, it responds, you talk again. GPT-Live processes continuously, making decisions multiple times per second about when to speak, when to pause, and when to stay quiet. The model can drop a quick "mhmm" to acknowledge you're still talking, or just wait. One HN commenter put it perfectly: "This solves my biggest annoyance with the current advanced voice, its speech getting interrupted by me setting up or even background noise if loud enough."

The architecture split is what makes this interesting. GPT-Live handles the conversation layer (listening, responding, managing turns) while delegating heavy computation to GPT-5.5 in the background. When you ask a complex question, GPT-Live keeps talking to you while GPT-5.5 does the actual reasoning. It's a two-tier system: fast conversational model up front, powerful reasoning model behind the curtain. OpenAI calls this "delegation for deeper work," and it's the same pattern they've been building toward with their agent frameworks.

There are two versions shipping today. GPT-Live-1 is the default for Go, Plus, and Pro subscribers. GPT-Live-1 mini powers the free tier. Both support three reasoning levels (Instant, Medium, or High) letting users trade speed for depth on the fly. The visual integration is new too: while you're talking, ChatGPT can now display cards for weather, sports scores, stocks, and maps. It's the first time voice mode feels like it has a visual channel, not just an audio one.

OpenAI isn't first to full-duplex. NVIDIA shipped PersonaPlex in January, a 7B open-weight model that does the same simultaneous listen-speak thing, with 18x lower latency than Gemini Live. Kyutai's Moshi has been doing full-duplex since late 2024. What GPT-Live adds is scale: it's rolling out globally to hundreds of millions of ChatGPT users today, across iOS, Android, and the web. The open-source models are impressive in benchmarks, but they require self-hosting and GPU access. GPT-Live just works in the app you already have.

The safety story is worth reading. OpenAI published a system card alongside the launch, detailing how they handle real-time voice risks. The system can steer responses mid-conversation, play spoken safety messages, or terminate calls in high-risk scenarios. They tested against production prompts (real anonymized user audio) and synthetic prompts (AI-generated edge cases). The results showed GPT-Live matching or exceeding the previous Advanced Voice Mode on safety metrics, with minor regressions in emotional reliance and sexual content categories. The delegation architecture complicates safety: GPT-Live inherits the safety training of whatever model it delegates to, which at launch is GPT-5.5.

One thing missing: video and screen sharing. GPT-Live doesn't support voice with video at launch. Users who need that can still access legacy ChatGPT Voice. It's a strange gap given how much OpenAI has invested in multimodal capabilities, and it suggests the full-duplex architecture is harder to extend to video than the turn-based one.

The pricing question matters for developers. GPT-Live isn't available via API yet; OpenAI has a sign-up form for notifications. When it arrives, it'll enter a market where Gemini Live charges around $0.005/min input and $0.018/min output, and AWS Nova 2 Sonic runs about $0.015/min. OpenAI's voice APIs have historically been more expensive than Google's, and the delegation to GPT-5.5 in the background suggests the cost structure will be higher than pure voice models. For now, ChatGPT subscribers get it included in their plan.

What GPT-Live actually changes is the interaction model. Voice AI has been stuck in a mode where the user adapts to the machine, pausing at the right moments, speaking clearly, avoiding interruptions. Full-duplex flips that. The machine adapts to the user. It listens while you think, waits when you hesitate, and only speaks when it has something to add. That's how human conversation works, and it's taken this long for AI to get there.

Sources