OpenAI adds standby mode to voice API, but the money-saving route may not be it
The GPT-Live-1.1 API is live, with more natural accents and standby support to cut costs. Developers have already started doing the math: leaving it idle for long periods may be worse than simply hanging up.
On November 19, OpenAI's GPT-Live-1.1 officially entered the API. It has been more than a month since it began powering ChatGPT Plus/Pro voice.
The new version has three core selling points: more natural cross-language accents, a new Standby mode to reduce idle costs, and more reliable task completion. The official demo video:
There's not much to say about the first point. 'Accent drift' in multilingual voice has always been a pain point: a model sounds normal in English, then suddenly sounds like a different person when switched to French. 1.1 claims to be more natural and more consistent; you'll need to test the actual effect yourself.
What's really worth doing the math on is Standby. This mode allows developers to set an idle timeout of 5–60 seconds. After the timeout, the model temporarily pauses inference, but the WebRTC connection remains open, and the rate drops from $0.05/minute to $0.02/minute. That looks 60% cheaper. The official video shows the process of entering standby after a conversation interruption and then waking up:
But for scenarios where no one speaks for a long time, one developer did the math: Standby is still billed by the minute, $0.02/minute, which is $1.2 per hour. If the gap between conversations lasts several minutes, it is often more cost-effective to simply close the session and restore from the saved state. In other words, Standby is suitable for 'temporarily leaving but coming back soon,' not for leaving it idle all day.
GPT-Live-1 also supports Standby, so existing users do not need to migrate.
Another signal: GPT-Live-1.1 has been in service in ChatGPT's voice mode since October 15, and only now is the API being opened. OpenAI's playbook is to validate in the product first, then give it to developers.
The community reaction is also interesting. Some asked whether the Spanish (Spain) accent had been fixed; someone built a fully local open-source version, OpenGPT-Voice, using Silero VAD + Faster-Whisper + a local LLM + Supertonic TTS to replicate the GPT-Live experience, supporting barge-in, with end-to-end latency of about 550–800 ms. Although it's just a prototype, it at least shows that a voice agent pipeline can already be assembled from open-source components:
Meanwhile, a group of users is protesting OpenAI's plan to retire Standard Voice Mode on December 11. More than a thousand people have signed on Change.org, and community posts say 'over three hours of reading time are demanding that the original voice be kept.' On one side, more advanced Live voice is moving forward; on the other, users want the old feature not to go away. Technological progress does not mean every decision will be popular.
The improvements GPT-Live-1.1 brings to voice agents are real: more natural speech, accurate understanding of letters and numbers, and the ability to delegate complex tasks to backend models without interrupting the conversation. But what developers care about more may not be just model capability, but also the cost structure and what users actually want.
Docs: https://developers.openai.com/api/docs/models/gpt-live-1.1
Open-source local version: https://github.com/vishalsingh2972/open-gpt-voice
发布时间: 2026-11-20 05:42