By OpenAILearning Objectives
- Understand what full-duplex voice means and why it is architecturally different from a chained voice pipeline
- Price a GPT-Live-1 deployment, including the part that is not the per-minute rate
- Decide when the Realtime API is the right surface and when batch text-to-speech or a specialist vendor is better
What Is the OpenAI Realtime API?
The OpenAI Realtime API is OpenAI's WebSocket-based interface for bidirectional, low-latency voice conversation — for builders who need a model to listen, reason and speak as a conversation unfolds, rather than processing text and audio in separate batch calls.
Since September 10, 2026 its headline model is GPT-Live-1, first introduced inside ChatGPT and now available to developers.
🎯Tip
Documentation: developers.openai.com/api/docs/guides/audio — WebSocket setup, session configuration and tool-use patterns.
Full duplex is the actual change
A conventional voice agent chains three components: speech-to-text, then a reasoning model, then text-to-speech. Each handoff adds latency, and each is a place to lose timing, context or the rhythm of a conversation. The result is the stop-start cadence that makes most voice bots feel like walkie-talkies.
GPT-Live-1 listens and speaks at the same time, reasoning over incoming and outgoing audio together in a single model. That is what allows it to handle an interruption gracefully, tolerate a pause without barging in, and ignore a cough or a background voice rather than treating it as a turn.
💡Key Concept
Why one model beats three. In a chained pipeline nothing knows whether the user has actually finished talking — speech-to-text only knows that sound stopped. So the system guesses with a silence timer, which is why these agents either interrupt you or leave dead air. A full-duplex model hears the incoming audio while producing its own, so "has this person finished their thought?" becomes something it can reason about rather than something a timer decides.
The most concrete published evidence for this is from the language-learning app Speak, which reports that GPT-Live-1 cut interruptions during thinking pauses by almost 80 percent against its previous turn-based system — giving learners more time to think before the tutor replies. That figure is the customer's own, published by OpenAI.
It delegates the thinking
GPT-Live-1 is a voice layer, not a frontier reasoning model. It delegates reasoning and tool calls to a backend text model — one of OpenAI's own, or a third-party model — and returns the answer in speech.
This is the design decision that most affects your bill and your latency budget, because you are running two models, not one. OpenAI's own published evaluations name the backend used for each result (Astra at medium effort for the harder banking and tool-use suites, Terra at low effort for others), which is a useful signal that backend choice materially moves the numbers.
It can also be wired directly to Codex, so a voice session can hand work to a coding agent and speak the result back.
Pricing
GPT-Live-1 costs five cents per minute for the front-end voice layer, billed on connected session time. That is only half the bill: the backend model that does the reasoning is billed separately at its own rates, and OpenAI publishes no combined figure, so a deployment using a frontier backend at high effort will cost substantially more than five cents a minute.
⚠️Warning
GPT-Live-1 is five cents per minute for the front-end voice layer. That figure does not include the backend model that does the reasoning, which is billed separately at its own rates. A deployment using a frontier backend at high effort will cost substantially more than five cents a minute, and OpenAI publishes no combined figure — you have to price both halves yourself.
- Full-duplex listening and speaking, interruption handling, telephony
- Billed on connected session time
- Available in the API today
- Reasoning and tool calls delegated from the voice layer
- Choose per product — effort level moves cost significantly
- Can be an OpenAI model or a third-party one
- Access to voices beyond the published set
- Eligibility and request process are not public
- Not self-serve
The right comparison is not Realtime against batch text-to-speech — those are different product categories. It is Realtime against ElevenLabs Conversational AI, Deepgram Voice Agents and Gemini Live, the field of low-latency voice-agent platforms.
Voices and reach
GPT-Live-1 ships twelve new voices — Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta and Cinder — broadening what was a small set of real-time voices across accents and dialects. OpenAI says voice and language options will keep expanding. Custom voices are gated behind a sales conversation.
Telephony support is included, so full-duplex agents can be deployed on phone calls — restaurant reservations, customer support — which is where turn-based awkwardness is least tolerable.
Named deployments published at launch:
| Customer | Deployment | Reported result |
|---|---|---|
| Yelp | Yelp Host and Hatch, answering reservation and food-order calls | Improved turn-taking and call-handling rates; callers speak fuller sentences |
| Speak | Live Tutor Lessons | Interruptions during thinking pauses cut by almost 80 percent |
| Fin (Intercom) | Voice customer support | Described as moving support from stop-start toward the flow of a phone call |
| Cognition | Devin, hands-free | Talking through an idea or handing off work away from the keyboard |
Every figure in that table is reported by the customer and published by OpenAI. None has been independently verified.
The prior generation
The May 7, 2026 launch put three models on this API, and they remain the reference point for what GPT-Live-1 consolidates:
| Model | What it does | Billing |
|---|---|---|
| GPT-Realtime-2 | Frontier-class voice reasoning and dialogue | Per-token (input + output) |
| GPT-Realtime-Translate | Real-time translation, 70+ input languages, 13 output languages | Per-minute |
| GPT-Realtime-Whisper | Live speech-to-text transcription | Per-minute |
The architectural difference is worth stating plainly: GPT-Realtime-2 was a strong turn-based model, while GPT-Live-1 changes the turn-taking itself. For translation and live transcription specifically, the specialist models above are still the ones built for those jobs.
Strengths
- Full duplex removes the walkie-talkie feel — interruptions, pauses and backchannels are handled by the model rather than by a silence timer
- One model instead of three handoffs, which removes both latency and the brittle failure points between speech-to-text, reasoning and speech synthesis
- Backend flexibility — pair it with an OpenAI frontier model or a third-party one, and tune effort level against cost
- Telephony out of the box, which is where conversational awkwardness costs the most
- Twelve new voices across accents and dialects, a real widening from the previous real-time set
- Published customer evidence, including a specific measured claim from Speak rather than only vendor assertion
- Composes with Codex, so a voice front end can hand real work to a coding agent
Limitations & Considerations
- The five-cents-per-minute headline is only the voice layer. The backend reasoning model is billed separately and OpenAI publishes no combined price, so the advertised rate understates a real deployment
- Every performance figure is customer-reported and vendor-published — none is independently replicated
- Custom voices are not self-serve, requiring a sales conversation with unpublished eligibility criteria
- Not a narration model. For long-form audiobook or lesson audio, batch text-to-speech still produces better single-voice quality
- Not a translation or transcription specialist — the May 2026 models remain the better fit for those jobs
- WebSocket complexity — bidirectional voice is harder to debug than request and response; build for retry, reconnect and graceful degradation
- Voice carries identity signal beyond text, so retention and data-handling policy deserves more care than a text deployment
Best Use Cases
| Use Case | Best surface |
|---|---|
| Phone support or reservations where interruptions matter | GPT-Live-1 |
| Voice tutor or conversational practice | GPT-Live-1 |
| Hands-free handoff to a coding agent | GPT-Live-1 wired to Codex |
| Live multilingual support | GPT-Realtime-Translate |
| Real-time meeting transcription or captioning | GPT-Realtime-Whisper |
| Long-form narration, audiobooks, lesson audio | Batch text-to-speech (tts-1-hd, gpt-4o-mini-tts) or ElevenLabs |
| Voice cloning specifically | ElevenLabs |
| Open-source or on-premises voice | Whisper open weights plus an open text-to-speech model |
Getting Started
- Confirm an OpenAI API account with billing configured at platform.openai.com
- Read the audio guide at developers.openai.com/api/docs/guides/audio, starting with the WebSocket session example
- Choose your backend model before you build. It sets both latency and the larger half of your cost, and effort level moves both
- Build a minimal agent first — expect a working echo agent in one to three hours, and considerably longer for production error handling
- Test with interruptions on purpose: talk over it, pause mid-sentence, let someone else speak nearby. That is the behavior you are paying for, and the only way to know it works for your users
- Price a real session end to end, adding the backend model to the five-cent voice rate, before committing to per-minute unit economics
- Deploy to telephony only after the interruption behavior holds on your own audio conditions
Key Takeaways
- The Realtime API is OpenAI's WebSocket voice surface, and since September 10, 2026 its headline model is GPT-Live-1
- Full duplex is the substantive advance — the model listens and speaks simultaneously, so interruptions and pauses are reasoned about rather than guessed at by a silence timer
- It delegates reasoning to a separate backend model, which means you run and pay for two models
- Five cents per minute covers the voice layer only; the backend is billed separately and there is no published combined price
- Speak reports interruptions during thinking pauses cut by almost 80 percent — the most concrete published figure, and still a customer's own
- Twelve new voices and telephony support ship with it; custom voices require contacting sales
- It is not a replacement for batch text-to-speech, translation or transcription — the May 2026 specialist models and long-form narration models still win those jobs
