OpenAI released GPT-Live-1.1 (gpt-live-1.1) in the Live API on November 19, 2026, priced at $0.05 per active minute and $0.02 per standby minute, billed per second. Those official duration rates match GPT-Live-1, and any backend model reasoning or tool calls run during a session are billed separately.
What Did OpenAI Release in the GPT-Live-1.1 API?
GPT-Live-1.1 is OpenAI’s full-duplex voice model for multilingual conversations, exposed in the API under the model ID gpt-live-1.1 on the /v1/live/sessions endpoint. OpenAI documented the API launch in its official API changelog on November 19, 2026, followed by an announcement from @OpenAIDevs on X at 20:00:32 UTC the same day. Before reaching API developers, GPT-Live-1.1 had already powered voice conversations in ChatGPT Plus and Pro starting October 15, 2026, according to OpenAI.
On the official GPT-Live-1.1 model page, OpenAI defines the release around simultaneous listening and speaking alongside backend delegation. Because GPT-Live-1.1 operates in full duplex, the voice frontend can listen to user interruptions while speaking an answer, accept audio and text inputs, and return audio and text outputs with streaming and function_calling support. Instead of handling heavy multi-step reasoning inside the voice loop itself, developers write plain-language instructions telling gpt-live-1.1 when to hand a task off to a backend model or tool workflow.
Alongside the new model snapshot, the November 19 changelog introduced a WebRTC standby configuration for the Live API. OpenAI highlighted lower idle costs in the launch post; standby applies to both gpt-live-1 and gpt-live-1.1, and the per-minute list rates are identical across the two models.
How Much Does GPT-Live-1.1 Cost?
According to the OpenAI API pricing page, gpt-live-1.1 costs $0.05 per minute for active session time and $0.02 per minute for WebRTC standby time, billed per second without rounding session duration up to the next whole minute. Those list prices match gpt-live-1, so switching model IDs is a quality upgrade rather than a price cut.
| Model ID | API Endpoint | Active Voice Rate | WebRTC Standby Rate | Billing Granularity | Backend & Tool Billing |
|---|---|---|---|---|---|
gpt-live-1.1 |
/v1/live/sessions |
$0.05 / minute | $0.02 / minute | Per second | Billed separately at backend model/tool rates |
gpt-live-1 |
/v1/live/sessions |
$0.05 / minute | $0.02 / minute | Per second | Billed separately at backend model/tool rates |
Total cost for a Live voice agent has two separate components: duration-based charges for the voice session and standard usage charges for whatever backend model and tools execute delegated tasks. When your voice frontend delegates reasoning to a separate text model, those token charges follow standard API rates; see our AI API pricing comparison and our breakdown of GPT-6.2 Astra and Sol for how backend reasoning tiers are structured.
Worked Voice Cost Example and the 15-Second WebRTC Credit
OpenAI’s voice latency and cost guide explains how active seconds, standby seconds, and session startup interact on an invoice:
- WebRTC initialization credit: Under default startup behavior, calling
POST /v1/live/sessionsto create a WebRTC session bills 15 seconds of voice duration while the connection initializes. Once the session starts running, OpenAI credits that 15-second charge against your running duration charges. Do not add an extra 15 seconds on top of a call’s actual length when estimating bills; as OpenAI’s documentation notes, a 90-second active session already includes the 15 seconds billed during initialization. - What counts as active time: Active seconds include user and model speech, silent gaps before your configured idle timeout expires, and time spent waiting on backend reasoning or tool execution while voice inference remains running.
- Worked 90-second session calculation: OpenAI’s cost guide gives an example call with 60 seconds of active conversation and 30 seconds on standby. Calculated from the official list rates, the voice session charge is
(1 × $0.05) + (0.5 × $0.02) = $0.06, plus any separate backend model and tool charges incurred during or between sessions.
When reviewing ways to reduce AI API costs, note how client mute controls behave: muting the user’s microphone input leaves the GPT-Live session active rather than pausing inference. When a conversation finishes completely, your client or server should close the session and collect final usage rather than relying on a local mute button.
What Changed in GPT-Live-1.1 vs GPT-Live-1?
Comparing GPT-Live-1.1 with GPT-Live-1 shows a voice-quality and multilingual upgrade on the same session architecture. When OpenAI brought GPT-Live-1 to general availability in the API on September 10, 2026, that changelog entry listed the active per-second voice billing model with backend usage charged separately, and made no mention of a standby rate. On November 19, 2026, OpenAI shipped gpt-live-1.1 alongside the new WebRTC standby configuration for both models.
OpenAI did not publish word-error-rate tables or acoustic benchmark scores on the checked documentation pages, so the documented differences come from OpenAI’s model pages and developer announcements:
- Multilingual transcription and pronunciation: OpenAI’s official changelog and model documentation state that
gpt-live-1.1transcribes speech more accurately thangpt-live-1and improves pronunciation alongside more consistent accent and voice delivery across languages. - Alphanumeric accuracy and task reliability: In the November 19
@OpenAIDevslaunch thread, OpenAI states thatgpt-live-1.1“completes tasks even more reliably,” follows instructions, and helps voice agents understand “precise details like letters and numbers” while delegating complex workflows to a chosen backend model. - Separate rate-limit pools: Both models use
/v1/live/sessions, yetgpt-live-1.1andgpt-live-1enforce separate concurrent-session limits, with each model’s limit applying across inference regions. OpenAI’sgpt-live-1.1model page directs developers to check their organization’s rate-limit settings rather than publishing fixed tier numbers on the page.
How Does GPT-Live Standby Work on WebRTC?
GPT-Live standby is a session configuration for gpt-live-1 and gpt-live-1.1 WebRTC connections that pauses model inference during silence while keeping the underlying call open. OpenAI documents standby exclusively for WebRTC sessions in its guide to managing Live conversations.
To enable standby when creating a WebRTC session via POST /v1/live/sessions, developers configure three optional behavior controls and listen for lifecycle events:
- Wait for speech on startup: Set
session.wake_policy.typetospeechso the session waits for someone to speak before starting active inference. - Configure the idle window: Set
session.idle_timeout_sto a whole number from5to60. After that many seconds pass without an input or output transcript, inference pauses and the session switches to standby billing. Omittingsession.idle_timeout_skeeps inference active after the first wake until the session ends. - Tune wake sensitivity if needed: Adjust
session.wake_policy.vad_thresholdbetween0and1(default0.5). Higher thresholds demand stronger evidence of human speech before waking the session, both at initial startup and after every standby pause. - Keep the mic track live and track events: Your client must keep the microphone audio track active during standby — muting the mic does not enter standby, and cutting the track prevents speech detection. The WebRTC connection and session ID stay unchanged, emitting
session.pausedwhen the idle timeout stops inference (while the connection and sideband stay open) andsession.resumedwhen speech wakes inference and restores conversation history.
One latency tradeoff accompanies standby: OpenAI notes that after a session pauses, the model’s first response to a short utterance can take longer while inference resumes.
Endpoints, Configuration Limits, and Unsupported APIs
GPT-Live-1.1 runs exclusively on the Live API (/v1/live/sessions, including POST /v1/live/sessions for session setup and wss://api.openai.com/v1/live/sessions for WebSocket signaling). Developers cannot pass gpt-live-1.1 to Chat Completions, Responses, the older Realtime API (/v1/realtime), Realtime transcription or translation, Assistants, Batch, Fine-tuning, Embeddings, Speech generation, or Transcription endpoints. All of those surfaces are marked unsupported on the gpt-live-1.1 model page.
Within a Live session configuration, two structural rules govern how you wire up your voice agent:
- Immutable model and voice per session: Both
model(set to the default and only listed snapshot,gpt-live-1.1) andaudio.output.voice(defaultmarin, alongside OpenAI’s built-in multilingual voices and custom voices) are locked once a session starts. Changing either requires starting a new session. - Instruction budget and delegation modes: System
instructionsaccept up to 16,384 tokens, where you describe in plain language when to delegate work. Settingdelegation.typeselects eitherresponsesmode orclientmode (leaving the field omitted ornulldefaults toclientmode).
Practical Chooser: Upgrading to 1.1, Standby vs Closing Sessions
Choosing how to configure gpt-live-1.1 and session lifecycle controls comes down to call patterns and organizational quota; for broader framework decisions across text and multimodal endpoints, see our guide on how to choose an AI model.
- When to switch your model ID to
gpt-live-1.1: Updatemodeltogpt-live-1.1if your application handles multilingual callers, switches languages mid-call, or collects spoken account IDs, confirmation codes, and spelled-out names where OpenAI targets stronger transcription and letter/number recognition. Since per-second duration rates matchgpt-live-1, testing the new snapshot adds no voice-rate markup. Before cutting production traffic over, verify your organization’sgpt-live-1.1concurrent-session limit in the OpenAI dashboard, because its quota pool is separate fromgpt-live-1. - When to enable WebRTC standby: Turn on
wake_policy.type: "speech"and a5to60secondidle_timeout_swhen users stay connected to a WebRTC voice interface while reading screen results, filling out forms, or pausing intermittently for half a minute to a couple of minutes. Standby keeps conversation history and the WebRTC transport warm at 60% less than the active voice rate during silent stretches. - When to close the session instead: Tear down the session and open a fresh one later whenever an application can cleanly disconnect during long gaps or when a user walks away for several minutes. An abandoned standby tab still accrues duration charges continuously, and your application state can coordinate backend work across separate sessions between calls.
FAQ
Is GPT-Live-1.1 cheaper than GPT-Live-1?
No. OpenAI lists identical duration rates for gpt-live-1.1 and gpt-live-1 ($0.05 per active minute and $0.02 per WebRTC standby minute, billed per second without whole-minute rounding). Any cost reduction comes from enabling WebRTC standby or closing idle sessions earlier rather than changing model IDs.
Does GPT-Live standby work on GPT-Live-1 too?
Yes. Although OpenAI announced WebRTC standby in the same November 19, 2026 changelog entry and X thread as gpt-live-1.1, standby is a Live API session configuration supported on both gpt-live-1 and gpt-live-1.1 with a 5 to 60 second idle_timeout_s on WebRTC connections.
Does muting the microphone trigger standby?
No. OpenAI’s voice cost documentation states that muting microphone input leaves the session active at the full active rate. WebRTC standby requires keeping the microphone audio track live so voice activity detection can detect when the user speaks again and resume inference.
Can I call gpt-live-1.1 on Chat Completions, Responses, or /v1/realtime?
No. The gpt-live-1.1 model ID is only supported on /v1/live/sessions. OpenAI marks Chat Completions, Responses, Realtime (/v1/realtime), Assistants, Batch, and standalone Transcription or Speech generation endpoints as unsupported; to use Responses for reasoning, configure delegation from the Live session.
Does the 15-second WebRTC initialization charge add extra cost to every call?
Not once the call runs for at least 15 seconds. Creating a WebRTC session via POST /v1/live/sessions bills 15 seconds during initialization under default startup behavior, and OpenAI credits that amount against running duration charges, so a 90-second active call bills for 90 seconds total.
Sources
- OpenAI API Changelog (November 19, 2026 entry)
- OpenAI Developers (@OpenAIDevs) announcement on X
- GPT-Live-1.1 Official Model Documentation
- GPT-Live-1 Official Model Documentation
- OpenAI API Pricing (GPT-Live Audio Table)
- Managing GPT-Live Sessions Guide (Standby & Idle Configuration)
- Voice Latency and Cost Optimization Guide



