On November 19, 2026, OpenAI's developer account announced that the voice model GPT-Live-1.1 is now available to developers through the API, under the model ID gpt-live-1.1. Launched alongside it is Standby billing: once a call goes silent beyond a set time, inference pauses while the connection stays open, at $0.02 per minute — 60% cheaper than the standard $0.05 per minute for active calls. The point of this update is not how much smarter the model got, but that keeping a voice agent online for long stretches has become cheaper on the bill for the first time.
Silence is the expensive part: how Standby billing works
The real cost of a voice call is often not in the talking: a support agent waits while a customer looks up an order number, the line goes quiet for a long time, and the model keeps billing by the minute. Standby mode targets exactly that empty stretch: developers can set the silence timeout to between 5 and 60 seconds; once it is exceeded, the model pauses inference while the WebRTC connection and conversation history stay intact, resuming automatically when someone speaks. Silent time is billed at $0.02 per minute, per second. By OpenAI's own math, ten minutes of silence costs $0.20 in Standby versus $0.50 at the standard rate. The option is available for both GPT-Live-1 and 1.1, and it covers only the voice session itself — backend model and tool usage are billed separately.
GPT-Live-1.1 is not a brand-new model: since October 15 it has powered voice in ChatGPT Plus and Pro, and this is the first time it is being handed to API developers. OpenAI lists three concrete improvements: more stable accents across languages, so speech in German or Japanese is less likely to drift toward an English accent; more accurate recognition of precise details such as spelled letters, phone numbers, order IDs and confirmation codes; and stronger instruction following, so complex work can be handed to a backend agent without interrupting the spoken conversation. The API also adds 38 voice IDs, seven of them new, including the Spanish-leaning calandria, hornero and luciernaga.
For developers, the trade-offs are equally clear: GPT-Live supports full-duplex conversation — listening while speaking — with WebRTC as the recommended connection, plus WebSockets and telephony SIP; but it does not support structured outputs or fine-tuning, its inputs and outputs are text and audio only, it cannot handle images or video, and there is no self-hosted option. In other words, it fits always-on voice entry points such as customer service and phone assistants; for text tasks that need fine control, another interface is the better choice. The voice race has been busy this year anyway — ElevenLabs just bought Japanese voice company Kotoba for over $100 million, and Microsoft's new voice model costs 1 cent per minute — the competition in voice models is shifting from 'can it talk' to 'what does it cost to stay on the line in silence.'