OpenAI's GPT-Live-1 Cuts Voice Agent Code by 80% at $0.05 a Minute

OpenAI just brought its full-duplex conversational voice model to the API. GPT-Live-1 is the same model that powers ChatGPT's natural back-and-forth voice mode, and it's now something you can wire into your own apps, phone systems, and agent stacks for $0.05 per minute on the voice layer.

Instead of gluing together speech-to-text, an LLM, and text-to-speech, you get one model that listens and speaks at the same time. Traditional voice agents chain those three components, and each handoff adds latency and creates opportunities to lose timing, context, or the natural rhythm of a conversation. GPT-Live-1 collapses that pipeline and lets your backend model do the heavy thinking asynchronously while the audio keeps flowing.

Why full-duplex changes the shape of a voice agent

Most voice agents today are turn-based: the user talks, the system waits for silence, then processes, then replies. GPT-Live-1 reasons over incoming and outgoing audio simultaneously, so it can handle interruptions, backchannels like mhm, and mid-thought corrections without breaking. Acknowledgements and course-corrections happen live, while deeper reasoning gets delegated to the back end.

Across OpenAI's evaluations, GPT-Live-1 improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1, with sizeable gains in turn-taking latency and interactive behavior. Paired with GPT-6 Astra at medium reasoning effort, it also ranks #1 on Tau3, which measures frontier voice-agent intelligence on end-to-end tasks.

The delegation pattern

The interesting architectural move: GPT-Live-1 isn't trying to be your reasoning brain. It hands off reasoning and tool calls to a backend text model like GPT-6 Astra or a third-party equivalent. You pick the model, the tools, and the agent harness that sit behind the voice layer.

That opens the door to mixing cheap-and-fast with smart-and-slow depending on the task. Developers might pair GPT-Live-1 with a model like Luna for high-volume work like scheduling or order updates, then route complex customer issues to something like Astra. The voice model keeps the caller engaged with acknowledgements and small talk while your backend chews through a tool call.

What you get out of the box

• Interruption handling that reasons over both sides of the audio stream at once
• Native ASR transcripts and response text, plus keyword biasing and strong alphanumeric understanding (useful for phone numbers, order IDs, addresses)
• Turn detection if you still want explicit turn boundaries in your app logic
• System-prompt control over tone, pacing, expressiveness, language, and response length
• Telephony support for deploying agents directly onto phone calls
• Background noise handling so cafe chatter or a second speaker in the room doesn't derail the session

Early customer numbers

A few launch partners shared concrete impact numbers rather than vibes. In early evaluations, Speak found that GPT-Live-1 gave learners more time to think before the language tutor responded, cutting interruptions by almost 80% versus previous turn-based systems. A healthcare customer reported that compared to their cascaded build, GPT-Live-1 shrank their code base by 80% and removed 23K lines of code, which is the more telling metric if you've ever tried to maintain a chained STT-LLM-TTS pipeline with custom interruption logic.

Yelp is using it for reservations and food-order calls, Cognition wired it into Devin so you can talk through code with the agent, and Fin is pushing it into customer support flows.

Voices, pricing, and getting started

The launch comes with twelve new voices spanning different accents and dialects (Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, Cinder), with more languages promised over the coming months. Custom voices require going through sales.

Pricing sits at $0.05 per minute for the front-end voice layer, on top of whatever your backend reasoning model costs. The API docs are live, and there's a reference pattern for wiring GPT-Live-1 to Codex so you can literally have a spoken conversation with an agent editing your repo in the background.

Where this leaves the voice-agent stack

If you've been building on cascaded pipelines, the math just changed. The main reasons to keep a custom chain were latency control, model choice, and being able to swap components. GPT-Live-1 keeps model choice (via backend delegation) and improves latency, which removes two of the three arguments. What remains is whether you trust a single closed model with the acoustic layer of your product. For phone-first products, tutoring, and support workflows where interruption handling is the actual UX, the tradeoff now looks pretty one-sided.