On September 10, 2026, OpenAI released GPT-Live-1 in the API, a voice model built for full-duplex conversation: it listens and speaks at the same time, handling interruptions and acknowledgments the way a person on a phone call does, rather than waiting for a speaker to finish before it responds. The model is priced at $0.05 per minute for the voice layer and ships with 12 voice options, support for custom voices, and built-in telephony connectivity so a voice agent can sit on an ordinary phone line. This piece walks through what changed architecturally, what it costs to actually run, and what the decoupled design signals about where voice agents are headed.
OpenAIFrom cascaded pipelines to native full duplex
Most voice assistants built over the past several years have used what is generally called a cascaded pipeline: speech-to-text converts the caller's audio into words, a text-based language model reasons over those words and drafts a reply, and text-to-speech turns that reply back into audio. Each stage is a separate model, run in sequence, and the caller waits for all three steps before hearing anything back.
That structure explains why so many voice bots still feel stilted. Turn-taking is rigid because the system cannot listen while it talks, latency stacks up because three models run one after another, and prosody, the rise and fall of natural speech, tends to get lost in the round trip through text. A cascaded system does not really hear a laugh, a half-finished sentence, or a "wait, actually" until the speaker stops entirely and the transcript catches up.
GPT-Live-1 targets that gap directly by treating voice as a native, continuous input and output rather than a text problem wrapped in audio. According to OpenAI, the model tracks the conversation stream in both directions simultaneously, which lets it register a caller talking over it, acknowledge a point mid-sentence, and yield the floor without the multi-second dead air that cascaded systems produce. OpenAI reports a 30 percentage point improvement on its internal Full Duplex Bench compared with GPT-Realtime-2.1, its prior real-time voice model. That figure is a vendor-reported benchmark rather than an independently verified result, and it should be read with that caveat, but the direction it points to, native full-duplex handling over cascaded turn-taking, matches the architectural description OpenAI has given for the release.
The architecture: a voice layer decoupled from reasoning
The more consequential design choice is not the full-duplex behavior itself but how OpenAI packaged it. GPT-Live-1 is explicitly scoped as the live conversation layer only. It is not a standalone assistant that reasons about a customer's account or looks up an order status; it handles the audio, the turn-taking, and the acknowledgments, then hands off the substantive thinking to a separate backend model that a developer chooses and pairs it with, such as GPT-6 Astra, Codex, or ChatGPT Work, along with whatever tools that backend calls.
This is a deliberate split between the perception and expression layer, which needs to be fast and continuous, and the reasoning layer, which can afford to be slower and heavier. In a cascaded system, one model (or one tightly bundled pipeline) had to do both jobs, which forced an awkward compromise: make the whole thing fast enough to feel conversational, at the cost of reasoning depth, or let it think longer, at the cost of feeling like a phone tree. Splitting the two lets each side optimize for what it is actually good at.
| Dimension | Cascaded pipeline (STT to LLM to TTS) | Native full duplex (GPT-Live-1) |
|---|---|---|
| Turn-taking | Rigid, waits for full stop | Continuous, handles interruption |
| Latency source | Three sequential model calls | One live audio-native pass |
| Reasoning backend | Usually bundled with the voice stage | Decoupled, swappable (GPT-6 Astra, Codex, ChatGPT Work, etc.) |
| Billing shape | Per-stage token and audio costs | Voice layer priced separately from backend usage |
The interesting design choice is not that the model can talk over itself, it is that talking is now billed and served separately from thinking.
The tradeoff cuts both ways. Composability is a real benefit: a developer can point the same voice layer at a cheap backend for routine calls and a more capable one for complex cases, or swap backends as better reasoning models ship, without re-architecting the voice stack. Set against that is added integration complexity, since a developer now has to wire two systems together and manage the handoff between them, rather than working against a single bundled assistant.
What a voice session actually costs
The decoupling also changes how buyers need to think about cost. The $0.05 per minute figure covers only the voice layer. Over a one-hour call, that is $3.00 for voice alone, before a single token of reasoning has run. Any backend model calls, tool use, or retrieval the conversation triggers are billed on top of that, using whatever the paired backend model charges.
That structure means the headline $0.05 per minute understates total cost for any agent that does real work during the call, and it means two developers running the same voice layer can pay very different totals depending on which backend, and how much tool use, they pair it with. It also means the per-minute voice price is the wrong number to optimize in isolation; the number that matters is cost per resolved call, which depends on how efficiently the backend reasons and how many tool calls a typical conversation needs.
Telephony support and where this is headed
OpenAI paired the release with built-in telephony support, letting a GPT-Live-1 agent connect directly to traditional phone networks rather than requiring a custom bridge. That detail is a fairly clear signal of intended use: call centers, receptionist and scheduling lines, and other phone-based workloads that have historically been served by either human agents or much more limited interactive voice response systems. Full-duplex handling matters disproportionately here, because a caller on a real phone line who gets interrupted or ignored mid-sentence is a far worse experience than the same lag in a chat window.
The same week, OpenAI also moved its Agents API into public beta, adding session management, tool coordination, and recovery mechanics aimed at longer-running agent workloads. GPT-Live-1 is the more voice-specific release of the two, but both point at the same broader push: turning agentic capability that has mostly lived in demos and chat interfaces into infrastructure developers can build production systems on.
None of this is happening in a vacuum. Other providers already offer real-time voice APIs, and the meaningful differentiators here are the full-duplex quality claim, the breadth of voice options, and the telephony integration, not the existence of a real-time voice product itself. Custom voices also raise the now-familiar questions around consent and impersonation that any voice-cloning capability invites, and reliability on real phone networks, with their variable audio quality and dropped calls, is a different engineering problem than reliability on a clean WebRTC connection in a demo.
For teams evaluating this, the decoupled voice-layer-plus-swappable-backend pattern is itself worth noting as a design principle, independent of which vendor ships it: keeping the reasoning brain interchangeable, rather than locked to whatever model shipped with the voice layer, is the same model-agnostic instinct Metir AI applies at the application layer, routing each task to whichever model handles it best rather than binding a workflow to one provider's stack. As voice and reasoning continue to be sold and priced separately, that flexibility is likely to matter more, not less.
Sources:
- Introducing GPT-Live-1 in the API | OpenAI
- OpenAI launches GPT-Live-1 for full-duplex voice agents | TestingCatalog
- OpenAI GPT-Live-1 API: Full Duplex Voice, Telephony, Pricing | DataStudios
- Full-Duplex Voice AI Hits Developer APIs: GPT-Live-1 Kills Cascaded Pipeline | Tech Times
- OpenAI Releases GPT-Live-1 in the API, Powering Natural Full-Duplex Voice Conversations, Telephony | AIToolly
Image credits
Header and in-body photograph: The Pioneer Building, San Francisco, OpenAI's headquarters, HaeB, via Wikimedia Commons, licensed under CC BY-SA 4.0.