GPT-Live-1.1 hits OpenAI's API: real-time voice AI for 5 cents a minute
OpenAI has opened its GPT-Live-1.1 voice model to developers through its API, with billing per active minute of conversation. The model can listen and talk at the same time. In ChatGPT, the GPT-Live family fully replaces the old voice modes on December 11.

What's happening
OpenAI has brought its voice model GPT-Live-1.1 to its API, the interface developers use to build OpenAI models into their own apps, phone hotlines and voice assistants. According to ad-hoc-news, GPT-Live-1.1 directly succeeds GPT-Live-1, which got its own system card on October 15, 2026. A system card is a technical document that describes a model's capabilities and safety testing.
The GPT-Live family has been around since July. That's when OpenAI rolled out GPT-Live-1 and its smaller sibling, GPT-Live-1 mini, in ChatGPT. The Next Web puts the date at July 8. What's new is that developers can now use the improved 1.1 version directly, with its own per-minute pricing.
What makes GPT-Live different
The models are full-duplex. Like a person on a phone call, they can listen and speak at the same time. Older voice modes were turn-based: they waited for you to finish, which often meant they cut in at the wrong moment. The even older "Standard Voice" mode chained three models together, according to an overview on note.com. One turned speech into text, one wrote the reply and one read it aloud. That made it slow.
According to eWeek, GPT-Live decides several times a second whether to speak, pause, keep listening, interrupt or use a tool. It can drop in short cues like "mhm" or "got it," wait quietly while you think, and stay silent until you address it again if you ask it to. In press demos, OpenAI showed off simultaneous translation, with the model translating while the speaker was still talking. TechRadar called it "genuinely useful," as The Next Web reports.
The second key feature is delegation. When a question needs a web search, deeper reasoning or agentic capabilities (completing multi-step tasks on its own), GPT-Live passes it to a more powerful text model in the background while the conversation keeps going. At launch in July, that model was GPT-5.5. ChatGPT Voice product lead Atty Eleti told CNET, as quoted by The Next Web:
That's exactly how people interact with each other. We keep the conversation going while we think in the background.
ChatGPT also lets you pick a reasoning level (Instant, Medium or High). For topics like weather, stocks and sports, the model shows visual cards.
API pricing and technical limits
For developers, the billing model is the biggest change. According to ad-hoc-news, the pricing works like this:
| Item | Price |
|---|---|
| Regular voice session | 5 cents per active minute |
| Qualifying WebRTC session with inference actively paused | 2 cents per minute |
| Backend reasoning (delegated tasks) | billed separately |
| External tool calls | billed separately |
WebRTC is a standard for real-time audio and video in the browser, the kind of thing video calls run on. Read the fine print: the cheaper rate only applies when inference, meaning the model's actual compute, is paused. Ordinary pauses in the conversation don't count. And because reasoning and tool calls cost extra, a minute can end up costing well over 5 cents, depending on your app.
On the technical side, GPT-Live-1.1 supports streaming, so responses arrive continuously rather than all at once, and function calling. That lets the model trigger things like a booking or a database lookup in your own software. Structured outputs (responses in a fixed data format) and fine-tuning (training the model further on your own data) aren't supported. Neither are images or video. GPT-Live-1 and GPT-Live-1.1 have separate quotas for concurrent sessions, and those quotas apply across regions.
Safety: better scores, with a caveat
In the accompanying system card, OpenAI says GPT-Live-1.1 scores better than its predecessor on mental health, emotional reliance and self-harm. On sexual content, illegal behavior and handling of personal data, OpenAI says it performs on par with GPT-Live-1. There are two important caveats. The safety tests were run with backend delegation turned off, so without the background model that's part of every real conversation. And OpenAI doesn't share actual rates of unsafe output in production.
This matters a lot for voice assistants. The Next Web notes that an AI that sounds like a friend wins trust more easily, which also makes it easier for it to mislead people. Signal president Meredith Whittaker has long warned that chatbots are "not your friends." OpenAI stresses that GPT-Live is meant to be an assistant, not a companion. The company says the model uses preset voices and doesn't imitate real people. Real-time safeguards can redirect a response, point users to crisis resources or end the conversation in risky situations, and extra protections apply to teens.
Weak spots and the switchover in ChatGPT
The system isn't flawless. In a live demo of translation into Hindi, eWeek reports, the assistant spoke with a heavy American accent and in stilted, "bookish" language. It's not known which languages OpenAI has optimized most. If you're building for languages other than English, test the quality yourself.
In ChatGPT, GPT-Live will soon be the only option. According to OpenAI's Help Center, summarized on note.com, "Advanced Voice" and "Standard Voice" shut down on December 11, 2026. Users will be moved to "Live Voice" automatically. OpenAI says voice usage has nearly doubled since GPT-Live launched. According to The Next Web, more than 150 million people use ChatGPT's voice and dictation features every week. Past voice conversations stay in your history. Live video and screen sharing will keep running on Advanced Voice for now, since GPT-Live can't do them yet. There's no date for when that changes. Some third parties already use GPT-Live, including the language-learning app Speak for new lessons.
According to ad-hoc-news, the API release comes shortly after a bigger shake-up in OpenAI's model lineup. On November 17, OpenAI reportedly introduced GPT-6.2 Astra for complex specialist work and GPT-6.2 Sol for everyday tasks, available in ChatGPT Plus, Pro, Enterprise and Edu and through the API. The sources don't say whether either model now handles GPT-Live's background tasks.
The bigger picture
With GPT-Live-1.1 in the API, OpenAI is opening up its full-duplex voice tech for call center bots, translation apps and voice control in third-party software. The 5-cent per-minute price is easy to budget for, but you only get the real cost once you add the separately billed reasoning and tool calls. Competition is fierce: ElevenLabs, startups like Sesame and Monogram, and Apple and Amazon with their upgraded assistants are all chasing the same market.
OpenAI is betting that talking, not typing, will become the main way people work with AI agents. It's still unclear how well that works outside English. The system card's safety scores were measured without the background model, and there are no numbers from real-world use, which deserves some skepticism. The Next Web also asks the obvious question: do people actually want an AI that never quite stops listening?
Related videos
Sources
- Bing News (de-DE) ad-hoc-news.de
- eweek.com eweek.com
- note.com note.com
- thenextweb.com thenextweb.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.