Last updated: November 21, 2026
OpenAI released GPT-Live-1.1 on November 19, 2026. The new gpt-live-1.1 voice model for the Live API improves pronunciation, accent consistency and transcription in multilingual conversations, and takes delegation rules in plain language. Pricing matches GPT-Live-1 at $0.05 per active minute, and new standby billing at $0.02 per minute cuts the cost of idle WebRTC sessions.
Contents
- What is new in GPT-Live-1.1
- GPT-Live-1.1 vs. GPT-Live-1
- Pricing and standby billing
- How to start or upgrade a session
- How to configure standby
- Prompt patterns for GPT-Live-1.1
- Limits and known constraints
- Frequently asked questions
What is new in GPT-Live-1.1
The OpenAI API changelog entry for November 19 lists two changes. First, GPT-Live-1.1 “improves pronunciation, accent consistency, and transcription in multilingual voice conversations.” Second, developers can configure standby for both GPT-Live-1 and GPT-Live-1.1 WebRTC sessions, which pauses inference during inactivity and resumes when someone speaks.
Like its predecessor, the model is full duplex: it can listen while it speaks. According to the GPT-Live-1.1 model page, it accepts audio and text input, produces audio and text output, and supports streaming and function calling.
The practical change for developers is in prompting. GPT-Live-1 used a special token to decide when to hand work to a backend. With GPT-Live-1.1, you describe in plain language when the model should delegate reasoning or tool use. Our guides to GPT-Live SIP outbound calling and GPT-Live voices and custom voices cover the shared session API.
GPT-Live-1.1 vs. GPT-Live-1
The table compares the two models using the GPT-Live-1.1 and GPT-Live-1 model pages and the prompting guide.
| Item | GPT-Live-1.1 | GPT-Live-1 |
|---|---|---|
| Model ID | gpt-live-1.1 |
gpt-live-1 |
| Multilingual quality | More accurate transcription; more consistent accent and voice across languages | Baseline |
| Delegation prompt | Plain language (“Delegate to the backend”) | Special token (Generate <|SpawnThinking|>) |
| Endpoint | v1/live/sessions only |
v1/live/sessions only |
| Active / standby price | $0.05 / $0.02 per minute | $0.05 / $0.02 per minute |
| Standby support | Yes (WebRTC) | Yes (WebRTC) |
| Concurrent-session limits | Separate from GPT-Live-1 | Build 50, Launch 300, Grow 500; Free tier unsupported |
OpenAI publishes no benchmark figures for the multilingual gains, so test with your users’ languages, names and accents before switching traffic.
Pricing and standby billing
The OpenAI pricing page lists the same rates for both models: $0.05 per minute while active and $0.02 per minute while on standby. GPT-Live bills voice sessions per second, and the GPT-Live-1 model page notes that duration is not rounded up to the next whole minute. Backend calls and tools are billed separately.
Two details from the latency and cost guide matter for estimates. Active time includes silence before the idle timeout expires and time spent waiting for backend work. Creating a WebRTC session with the default startup behavior bills 15 seconds while the session initializes, which is credited against duration once the session runs.
At the published rates, a fully active 10-minute session costs $0.50 in voice duration. With 6 of those minutes on standby, it costs $0.32 (4 minutes at $0.05 plus 6 at $0.02). Muting the microphone does not trigger standby.
For wider context on API costs, see our OpenAI API pricing overview.
How to start or upgrade a session
OpenAI’s getting started guide recommends WebRTC for browser, mobile and server audio, with WebSockets and SIP as alternatives.
- Set the model. In your session configuration, set
modeltogpt-live-1.1. It cannot change mid-session. - Choose a delegation mode. Set
delegation.typetoresponsesto let OpenAI call a Responses model you choose, orclientto run your own agent. Omitting delegation selects client mode. - Write instructions. Put conversation behavior in
instructions, up to 16,384 tokens, using the plain-language delegation template below. - Pick a voice. Set
audio.output.voice; the default ismarin. - Connect and wait. Your server creates the session through
POST /v1/live/sessions, the browser connects, and you wait forsession.startedbefore speaking. - Close cleanly. Send
session.closeand wait forsession.closedto collect final usage.
A minimal Responses delegation configuration, adapted from the delegation guide with the new model ID, looks like this:
{
"model": "gpt-live-1.1",
"delegation": {
"type": "responses",
"responses": {
"model": "gpt-6-luna",
"instructions": "[Your backend prompt]"
}
}
}
Use client delegation when you must validate or redact backend results before the caller hears them. For backend model choices, see our GPT-6.2 Astra and Sol guide.
How to configure standby
Standby is set when you create a WebRTC session, according to the session management guide:
- Set
session.wake_policy.typetospeechso inference waits for speech. - Set
session.idle_timeout_sto a whole number from 5 to 60. After that many seconds without an input or output transcript, the session returns to standby. If you omit it, inference stays active after the first wake. - Optionally set
session.wake_policy.vad_thresholdfrom 0 to 1 (default 0.5). Higher values need stronger evidence of speech. - Listen for
session.resumed(inference ready) andsession.paused(standby entered).
Keep the microphone track active during standby. The connection and session ID stay the same, and conversation history is restored when the user speaks. OpenAI notes that the first response to a short utterance after standby may take longer. For long gaps, OpenAI suggests closing the session and starting a new one with saved history.
Prompt patterns for GPT-Live-1.1
The Prompting GPT-Live guide provides a template for gpt-live-1.1 and asks you to keep three headings: Backchannel policy, Interruption policy and Delegation policy. A condensed version:
You are [name], a calm, friendly voice assistant for [service].
Backchannel policy: Use moderate backchannels.
Interruption policy: Stop speaking when the user interrupts.
Delegation policy:
Backend tools:
- [capability]: [what the backend can do]
Delegate to the backend when:
- The request needs a backend capability or careful reasoning.
Do not delegate when:
- You can answer from the conversation or a still-current result.
Do not guess the result while waiting.
For multilingual apps, the area 1.1 targets, the guide recommends these settings:
- Write the prompt and examples in the language you want the assistant to speak, and spell out pronunciations that matter.
- Add a line such as “Speak [language] unless the user asks to switch.”
- To open in a chosen language, put greeting instructions in a
developermessage insession.input. - Use
session.thinking.appendfor silent context, and summarize JSON or tables first, since large updates can increase hallucinations.
For consumer speech-to-text rather than API voice agents, see our ChatGPT Live Dictation guide.
Limits and known constraints
- Endpoint: only
v1/live/sessions; no Responses, Realtime or Batch access. - Context: the default context window holds 128,000 tokens, including instructions, text and audio tokens. Above 90% usage, GPT-Live starts a replacement engine with your original instructions and up to 8,192 tokens of history, so keep key facts in your application.
- Fixed settings: model, voice, delegation mode and storage are chosen at startup.
- Standby scope: standby applies to WebRTC sessions only.
- Rate limits: separate concurrent-session limits for 1.1 and 1, applied across inference regions.
Frequently asked questions
What is GPT-Live-1.1?
GPT-Live-1.1 is OpenAI’s full-duplex voice model for the Live API, released on November 19, 2026. It listens while it speaks and, compared with GPT-Live-1, transcribes speech more accurately and keeps a more consistent accent and voice across languages. The model ID is gpt-live-1.1.
How much does GPT-Live-1.1 cost?
OpenAI lists $0.05 per minute of active session time and $0.02 per minute of standby time, the same rates as GPT-Live-1. Sessions are billed per second, and backend model and tool usage is billed separately.
What is standby in GPT-Live?
Standby pauses inference in a WebRTC session when nobody is talking. Set session.wake_policy.type to speech and session.idle_timeout_s to 5 to 60 seconds. The connection and session ID stay the same, and the lower standby rate applies while inference is paused.
How do I upgrade from GPT-Live-1 to GPT-Live-1.1?
Change the model field in your Live session configuration to gpt-live-1.1. Then rewrite the delegation section of your prompt in plain language, because GPT-Live-1 prompts use the special SpawnThinking token instead.
Can I use GPT-Live-1.1 with the Responses or Realtime API?
No. The model page lists only the Live endpoint, v1/live/sessions, as supported. Chat Completions, Responses and Realtime endpoints are listed as not supported.
Sources
- OpenAI: GPT-Live-1.1 model page
- OpenAI API changelog (November 19, 2026 entry)
- OpenAI: Managing GPT-Live sessions
- OpenAI: Prompting GPT-Live
- OpenAI API pricing
Image: OpenAI logo (symbol), public domain, via Wikimedia Commons. OpenAI and GPT-Live are trademarks of OpenAI. FreeAnswers is not affiliated with OpenAI.