OpenAI Developers says GPT-Live-1.1 is now available in the API, in a post on X. The post claims three improvements: steadier accents across languages, a standby mode that lowers the cost of idle sessions, and more dependable task completion.

Original post on X

Load the post to view it as published on X. X may receive connection data.

View the original post on X ↗

The model is not new to everyone. The same post says it has powered voice in ChatGPT Plus and Pro since October 15, and FelloAI's explainer dates the upgrade to that day. The API release lets developers build it into their own voice agents.

What GPT-Live-1.1 does and how it differs from GPT-Live-1

The model page lists it as a full-duplex voice model, meaning it can keep listening while it talks. TechCrunch's report on the July launch explains that this design lets users interrupt naturally. It accepts audio and text and returns audio and text; image and video are not supported.

Compared with GPT-Live-1, the docs report more accurate transcription and a voice that stays more consistent across languages. The post adds that the model can reliably understand precise details like letters and numbers. In our view, that claim is the easiest to test, since phone numbers, order codes and spelled names are details a developer can check quickly.

How standby mode cuts the cost of silence

Standby lets developers set an idle window of 5 to 60 seconds. Once that window passes, inference pauses while the WebRTC connection stays open, and the model page's pricing puts that time at $0.02 per minute instead of $0.05.

Original post on X

Load the post to view it as published on X. X may receive connection data.

View the original post on X ↗

Media source: x.com

The post says standby also works with GPT-Live-1, but the pricing shown is for GPT-Live-1.1, so confirm the GPT-Live-1 rate before you budget. Take a ten-minute session with six minutes of pause. At the active rate it costs $0.50; with the six minutes at standby it costs $0.32 (4 × $0.05 + 6 × $0.02). That estimate is simplified: the idle window before the pause is likely billed at the active rate, and backend charges come on top.

How to start building with GPT-Live-1.1

Requests use the model ID `gpt-live-1.1` and go to the Live endpoint at `v1/live/sessions`. Streaming and function calling are supported; structured outputs and fine-tuning are not. Concurrent-session limits for GPT-Live-1.1 are separate from those for GPT-Live-1, so check your organization's limits before you plan capacity.

Delegation is the design choice to make first. Ordinary-language rules tell the model which requests should go to your backend for heavier reasoning or tool calls, and the post says you can pick that backend model. Those backend charges are billed separately from the voice rate, so the per-minute voice price is a floor, not your full bill.

What to check before you rely on it

The post reports no benchmark figures, and the model page lists performance and speed as "Not specified." The accent and reliability gains are OpenAI's own claims, so test with your languages, accents and background noise before switching a production agent. A short pilot with standby enabled will show both the silence bill and how accurately spelled-out details come through.

Key takeaways

  • GPT-Live-1.1 is available in the API as `gpt-live-1.1`, billed at $0.05 per minute while active.

  • Standby drops the paused-session rate to $0.02 per minute, and the post says it also works with GPT-Live-1.

  • The accent, reliability and speech-detail gains are OpenAI's own claims without published benchmarks, so test before switching.

Frequently asked questions

How much does GPT-Live-1.1 cost?

It costs $0.05 per minute while active and $0.02 per minute during standby. Backend model and tool usage are billed on top.

Can I use standby with GPT-Live-1?

The post says standby is available for GPT-Live-1 as well as GPT-Live-1.1. The pricing on the model page is for GPT-Live-1.1, so check the GPT-Live-1 rate first.

Does GPT-Live-1.1 handle images or support fine-tuning?

No. It takes audio and text, and the model page marks image, video and fine-tuning as not supported.

Sources