OpenAI Releases GPT-Live-1.1 Voice Model and Cuts Idle Session Cost

OpenAI has released GPT-Live-1.1, an updated full-duplex voice model available in its API, alongside a cheaper billing state for idle sessions. The company says the update improves accents and task completion, though it has not published benchmark scores to support those claims.
Model improvements and what OpenAI claims
According to the OpenAI team, GPT-Live-1.1 speaks with more natural and consistent accents across languages and completes tasks more reliably. The model is a full-duplex system for multilingual conversations, meaning it hears and speaks as a single stream rather than in turns.
OpenAI also describes better transcription and better handling of letters and numbers. These gains are self-reported, and the company has published no benchmark figures for version 1.1.
Standby mode lowers idle call cost
The release adds standby, a cheaper billing state for sessions that are idle. In standby, inference pauses during silence while the WebRTC connection stays open, and idle time is billed at $0.02 per minute instead of $0.05. For teams running voice agents, this reduces the cost of calls where the agent is waiting.
Standby can be configured through three settings: a wake policy set to speech, an idle timeout between 5 and 60 seconds, and an optional voice activity threshold between 0 and 1, defaulting to 0.5. After the timeout the server emits a session.paused event, while the session ID and history are kept. OpenAI warns that the first reply after a short utterance may be slower. Standby works for both GPT-Live-1 and GPT-Live-1.1.
Architecture and availability
The GPT-Live family replaces the usual chain of speech-to-text, a language model and text-to-speech, according to OpenAI, which says the model reasons over incoming and outgoing audio together. It also handles turn detection and returns transcripts natively, and can hand reasoning and tool calls to a backend text model.
GPT-Live-1.1 follows GPT-Live-1, which reached the API on September 10, 2026. As per an OpenAI forum announcement, the earlier model has powered ChatGPT Plus and Pro voice since October 15. Inputs and outputs are audio and text, while image, video, structured outputs, fine-tuning and predicted outputs are not supported. Live sessions hold a default context window of 128,000 tokens and run only on OpenAI's cloud via the v1/live/sessions endpoint.
Voice sessions are billed by duration, at $0.05 per minute or $3.00 per hour when active. Backend model and tool usage are billed separately.
Open questions
- How much better does GPT-Live-1.1 perform than GPT-Live-1 on independent measures?
- How much slower is the first reply after a short utterance in standby mode?