OpenAI opened its voice model GPT-Live-1 to API developers on September 10. The model first shipped inside ChatGPT's voice mode two months earlier, when the company said only that an API version would follow "soon." Developers can now build it directly into products such as phone support lines, voice assistants and spoken-language practice apps.
GPT-Live-1 is a full-duplex model: it keeps listening while the user is talking and keeps listening while it is talking itself, so it can be interrupted at any point, and it is built to handle background noise and long calls. Unlike the previous Realtime line, it isn't meant to work everything out on its own. When a request needs reasoning or a tool call, it hands the task off to whatever backend model the developer has configured, while it keeps the conversation flowing.
Front end and back end are billed separately
Pricing follows that same split. The voice front end costs $0.05 per minute, billed by the second rather than rounded up to the next full minute; whatever backend model or tool handles the reasoning is billed separately at its normal rate.
OpenAI suggests three pairings: Luna for high-frequency, simple tasks like scheduling or placing orders; GPT-6 Astra for support questions that need real reasoning; and Codex for anything code-related. Third-party models can also be plugged in on the backend.
According to the developer documentation, the model is called gpt-live-1 and runs on a new Live endpoint (v1/live/sessions) rather than the existing Realtime API. Input and output support audio and text but not images or video, and its knowledge cutoff is July 31, 2025. The concurrent-session limit ranges from 25 to 500, depending on account tier.
How it compares with GPT-Realtime-2.1
In the benchmark numbers OpenAI published, the biggest gap is on task-completion metrics:
| Metric | GPT-Live-1 | GPT-Realtime-2.1 |
|---|---|---|
| Tau3 voice intelligence (one-shot success) | 86.2% | 45.7% |
| Full-duplex interactivity | 80.1% | 45.4% |
| Tool-call one-shot success | 87.0% | 60.0% |
| Turn-taking latency | 0.798s | 1.41s |
| Artificial Analysis conversational dynamics | 97.3% | 95.7% |
On conversational naturalness, the two generations are only 1.6 points apart — the previous model was already close to a perfect score. The real gap opens up on "getting things done while talking": both Tau3 and the tool-calling benchmark require the model to carry a real business task through to completion during a live voice conversation. Architecturally, most of the gain comes from handing reasoning off to the backend while the front end sticks to listening and speaking. The Tau3 score was measured with a backend model attached — OpenAI's materials point to the GPT-6 Astra pairing — and the company hasn't published a separate score for what a cheaper backend like Luna would achieve.
What early customers are saying
OpenAI named Yelp, Speak, Fin and Cognition as early customers. Yelp CTO Alex Levy noticed that callers were starting to speak in fuller sentences:
"Callers are also speaking fuller, more natural sentences, which tells us the experience on the other end of the phone feels genuinely different."
Andrew Hsu, CTO of language-learning app Speak, put a number on it: the model interrupts users roughly 80% less often when they pause to think, compared with older turn-based systems. Jordan Neil, COO of support company Fin, described voice support moving from a stop-and-start back-and-forth to something closer to the rhythm of a normal phone call. Cognition connected the model to its coding agent Devin, letting engineers talk through a plan with the agent out loud while coding.
Generated audio carries a SynthID watermark. Only a small set of voices is available for now; custom voices require contacting sales.
What OpenAI hasn't disclosed this time is how the model performs in Mandarin. The company says language and accent support will keep expanding over the coming months but hasn't published a list of which languages are next. Teams building Mandarin phone support will need to run their own tests before deciding whether to switch.
Sources: OpenAI's official announcement and developer documentation for pricing, API and concurrency details, Unite.AI, CocoLoop, and public statements from customers including Yelp and Speak; all benchmark scores are as published by OpenAI.