OpenAI brings its ChatGPT voice model to the API with cheaper idle time, steadier accents, and better handling of letters and numbers.
-
GPT-Live-1.1 is now in the OpenAI API , same model that powers ChatGPT voice since October 15th.
-
New Standby mode drops idle billing from $0.05 to $0.02 per minute during silence.
-
Idle timeout is configurable between 5 and 60 seconds while keeping the WebRTC call open.
-
Improved accent consistency across languages and better handling of letters and numbers.
-
More reliable instruction following and backend delegation without breaking conversational flow.
-
Standby is available for both GPT-Live-1 and GPT-Live-1.1 on the v1/live/sessions endpoint. OpenAI has released GPT-Live-1.1 through its API, giving voice-agent developers more consistent accents, better handling of precise details, and stronger instruction following. A new Standby mode reduces the idle-session rate by 60%, from $0.05 to $0.02 per minute.
ChatGPT’s voice model moves into apps #
GPT-Live is a full-duplex voice model that can receive and generate audio simultaneously. It can handle interruptions and keep a conversation moving while a separate backend processes tool calls, web searches, or longer reasoning tasks.
OpenAI launched GPT-Live-1 on September 10, 2026, under the model ID gpt-live-1 and the /v1/live/sessions endpoint. ChatGPT Voice has used the 1.1 model family since October 15, with GPT-Live-1.1 or 1.1 Quip assigned according to the user’s plan. Developers can now access GPT-Live-1.1 directly.
Three upgrades aimed at real calls #
- More consistent multilingual accents: The model is less likely to drift toward an English accent while speaking languages such as German or Japanese.
- Higher accuracy for exact details: OpenAI says the model handles spelled letters, phone numbers, order IDs, and confirmation codes more reliably.
- Stronger instruction following: GPT-Live-1.1 can delegate complex workflows to a backend agent while maintaining the spoken conversation.
Silence now costs 60% less #
Standby mode lets developers configure an idle timeout from 5 to 60 seconds. After that period of silence, the model s while the WebRTC connection remains open. The session returns to the active rate when conversation resumes.
Ten idle minutes cost $0.20 in Standby, compared with $0.50 at the active rate. The feature is available for both GPT-Live-1 and GPT-Live-1.1.
| GPT-Live session rates | ||
|---|---|---|
| State | Rate | When it applies |
| --- | --- | --- |
| Active | $0.05 per minute | During conversation and before the idle timeout |
| Standby | $0.02 per minute | After 5 to 60 seconds of silence |
Keep speech fast, send heavy work elsewhere #
A typical deployment keeps the microphone and WebRTC connection in the client while the application server creates the GPT-Live session. OpenAI’s Client Delegation feature routes complex work to a separate agent without closing the voice session.
- The browser or mobile client captures audio and maintains the WebRTC connection.
- The application server creates and manages the GPT-Live session.
- GPT-Live handles low-latency listening and speech generation.
- A backend model uses the Responses API for web search, reasoning, and function calls.
- The backend returns results to the voice layer, which continues the conversation.
GPT-Live takes a separate API path #
GPT-Live builds on OpenAI’s earlier Realtime work but uses a separate endpoint and model lifecycle. Existing Realtime agents continue to use /v1/realtime, while GPT-Live sessions use /v1/live/sessions.
OpenAI’s deprecation schedule directs migrations from gpt-realtime to gpt-realtime-2.1 before January 20, 2027. Moving an existing integration to GPT-Live requires adapting session creation and transport in addition to changing the model ID.
Where the economics improve #
Drive-through ordering, in-car assistants, telehealth check-ins, and support lines often combine long s with exact details such as phone numbers and order IDs. Standby lowers the cost of keeping those sessions open, while the model update targets errors that can prevent a call from completing successfully.
Cloud and context limits remain #
- Cloud hosting only: GPT-Live has no self-hosted or on-premises option. That may prevent adoption where policy requires voice data to remain inside an organization’s infrastructure.
- OpenAI controls the voice layer: Client Delegation supports a custom backend agent, but OpenAI’s model continues to handle listening and speech generation.
- Audio consumes context: The 128,000-token context window includes audio tokens that may be absent from the transcript. Long sessions therefore need server-side state management, summaries, or conversation compaction.
Checks to run before shipping #
- Measure accent stability in each supported language, including code-switching between languages.
- Test phone numbers, names, confirmation codes, and order IDs under realistic noise conditions.
- Evaluate interruptions, overlapping speech, and latency during backend tool calls.
- Verify Standby entry and recovery across browsers, mobile networks, and WebRTC reconnections.
- Track context growth during long calls, including audio that does not appear in transcripts.
- Review data residency, retention, and access requirements before sending production audio.
Production evaluations should compare task completion, exact-string accuracy, interruption handling, and total session cost against the current voice stack. Those measurements will show how GPT-Live-1.1 and Standby perform under real audio conditions.