OpenAI dropped GPT-Live-1 into the API on September 10. It is a full-duplex voice model — it listens and speaks simultaneously — and the voice layer costs $0.05 per minute. That is the number developers have been waiting for. The harder question is what you pay for the backend reasoning that runs behind it, because the headline rate covers only one layer of a two-layer bill. Here is what the total actually looks like, and what it takes to build on this thing.
What Full-Duplex Actually Means #
Every voice agent built before GPT-Live-1 ran the same pipeline: speech-to-text captures the audio, a reasoning model processes it as text, a text-to-speech model reads the answer back. Three sequential models, each adding latency, each failing in its own way when the caller tries to interrupt. Interruption handling especially was a mess — developers wrote custom logic to detect the overlap, cancel the TTS buffer, and restart the reasoning pass.
GPT-Live-1 removes that pipeline. It is a single model handling audio input and audio output at the same time, making decisions many times per second about whether to speak, keep listening, , or hand the problem to a backend. Interruptions work the way they do in human conversation because the model has not pre-committed to finishing a sentence. It also delegates complex reasoning asynchronously — while it keeps the conversation alive on the phone, a backend GPT-4o or o3 instance does the actual thinking and returns results. No dead air.
The Real Price Tag #
The $0.05 per minute rate is real. It covers GPT-Live-1 itself, billed per second. What it does not cover is the backend model doing the reasoning, the tool calls, or your orchestration infrastructure. The total cost has three layers:
- Voice layer: $0.05/min (GPT-Live-1)
- Reasoning layer: backend model tokens billed at that model’s own rate
- Tool layer: function calls, database lookups, external APIs
In practice: a simple scheduling or FAQ bot with a lightweight backend runs roughly $0.07–0.08 per minute total. Add GPT-4o-class reasoning and you are at $0.12–0.18. Drop in o3 with tool calls and you are past $0.25. At 1,000 calls per day averaging five minutes, that is $400 to $750 per day — $12,000 to $22,500 monthly. Model choice is your biggest cost lever, not the voice layer itself.
This Is Not a Realtime API Drop-In #
If you are currently running the Realtime API with gpt-realtime-2.1, do not swap the model name and call it done. GPT-Live-1 runs on a different endpoint (Live sessions, not Realtime), uses a different event schema, and has restructured prompt handling. OpenAI publishes a migration guide, and the short version is that your prompts split into two buckets: session instructions for conversation style and delegation behavior, and delegation instructions for business logic and tool definitions. Manual turn control is also gone — the model decides when to speak. Your call-handling code will need adjustments.
The upside is that what you are migrating to is cleaner. Here is what a session start looks like on the new endpoint:
{
"type": "session.start",
"model": "gpt-live-1",
"instructions": "You are a customer support agent for Acme Corp. Keep responses brief.",
"audio": {
"output": { "voice": "quartz" }
},
"delegation": { "type": "client" }
}
Voice, speaking speed, and tone are all configurable through the instructions field. Audio format has to be set at session start and cannot change mid-call — decide upfront whether you are going WebSocket (server-side) or WebRTC (browser).
Where It Fits in the Landscape #
The old production voice stack was: Deepgram for ASR, GPT-4o or Claude for reasoning, ElevenLabs or Cartesia for TTS, Vapi or Twilio for orchestration. Four vendors, four bills, four failure points. GPT-Live-1 collapses the first and last layers — voice in, voice out — while keeping reasoning pluggable. Vapi, Retell, and ElevenLabs are still relevant as platforms that orchestrate on top of it.
The closest architectural peers are Gemini Live and Amazon Nova Sonic, both single-model speech-to-speech systems. Gemini 3.8 Live benchmarked ahead of GPT-Live-1 on some metrics at 40% lower cost. The competition is real and moving fast. GPT-Live-1’s advantage right now is the OpenAI ecosystem: if your reasoning layer is already GPT-4o or o3, the integration is straightforward.
Who Should Build on This Now #
EliseAI, an AI communications platform for homes and healthcare, was an early design partner and has GPT-Live-1 running in production telephony. That is the clearest signal of where this makes immediate sense: high-volume phone call workflows where natural interruption handling has real business value. Customer support lines, appointment scheduling, intake calls — anything where callers expect to be able to speak naturally rather than wait for prompts.
If you are building a new voice agent, start with the official docs and price out all three cost layers before you commit. The voice is good. The bill is manageable. The migration from Realtime is real work. Know what you are signing up for.