OpenAI released GPT-Live-1.1 on November 19, 2026, introducing a named model version for multilingual voice applications. The company says it improves pronunciation, accent consistency, and transcription, and developers can select it using gpt-live-1.1 in an existing Live session configuration. The same API changelog entry introduces configurable standby for GPT-Live-1 and GPT-Live-1.1 WebRTC sessions.
For builders, those are two changes worth testing separately. The model upgrade concerns conversation quality. Standby pauses inference during inactivity and resumes when someone speaks, with a lower per-minute rate while inference is paused.
That distinction matters: developers can evaluate standby without upgrading from GPT-Live-1, or test the new model before changing their session behavior. I’d separate those experiments initially, because otherwise a change in responsiveness or cost becomes harder to attribute.
What Changes With GPT-Live-1.1
GPT-Live-1.1 remains a full-duplex voice model, meaning it can listen while it speaks. It accepts audio and text inputs and produces audio and text outputs. OpenAI’s model documentation specifically compares it with GPT-Live-1, claiming more accurate transcription and a more consistent accent and voice across languages.
The practical targets are easy to recognize in a voice application:
- Pronunciation: Whether names, places, product terms, and ordinary speech sound correct in the intended language.
- Accent consistency: Whether the selected voice maintains its intended characteristics through a conversation, including language changes.
- Transcription: Whether spoken words are captured accurately enough to preserve the user’s meaning.
These are related, but they need separate checks. An assistant can produce understandable speech while misreading a customer’s name. It can also sound natural while transcribing an address incorrectly, potentially sending the wrong information into a downstream workflow.
OpenAI’s changelog provides no quantified multilingual comparison, and the opened sources contain no independent testing. The improvements are therefore vendor claims to validate against an application’s own languages and tasks, without assuming equal gains everywhere.
Model and API Compatibility
The model page documents a specific integration boundary:
| Detail | GPT-Live-1.1 |
|---|---|
| Model identifier | gpt-live-1.1 |
| Supported endpoint | v1/live/sessions |
| Input | Audio and text |
| Output | Audio and text |
| Conversation mode | Full-duplex |
| Listed features | Streaming and function calling |
| Session billing | Duration-based |
| Backend models and tools | Billed separately |
The endpoint is important. OpenAI lists the Responses API, Chat Completions API, and Realtime API as unsupported for this model. A developer with an existing Realtime integration should not read “live voice model” as permission to substitute this identifier into a Realtime request.
For an existing Live integration, the documented starting point is simpler: select gpt-live-1.1 in the session configuration. That is a model-selection change, but the small size of the edit says little about how much regression testing a production voice application needs.
A Practical Migration and Voice-Testing Plan
OpenAI also tells developers to describe in plain language when the voice model should delegate work to their backend. Its documentation positions GPT-Live-1.1 as the conversational interface, with backend delegation available for reasoning or tool use. That division deserves attention during migration, alongside speech quality.
A useful rollout sequence is:
- Change the model in a test configuration. Keep the existing instructions, backend, and other session settings stable for the first comparison.
- Replay the same multilingual scenarios. Compare GPT-Live-1 and GPT-Live-1.1 using matching inputs and evaluation criteria.
- Inspect speech and transcription separately. Listen for pronunciation and voice consistency, then check the recognized content.
- Exercise backend handoffs. Confirm that the model delegates the intended tasks and passes along the correct information.
- Test standby independently. Introduce inactivity and resumed speech after establishing the new model’s baseline behavior.
- Check concurrency limits before expanding traffic. The two versions have separate concurrent-session limits.
Those limits apply across inference regions for each model, according to the model documentation. A staged rollout therefore needs capacity checks for both the new model and any GPT-Live-1 fallback. Moving traffic between regions should not be treated as a way around a model-wide limit.
Test Language Changes Inside Conversations
A multilingual test set should include more than separate, single-language recordings. For applications whose users switch languages, include those switches within a conversation, along with borrowed words, local place names, abbreviations, numbers, and domain-specific vocabulary.
A useful test might ask an assistant to discuss a booking in one language, repeat a passenger’s name from another, and confirm an alphanumeric reservation code. Score the individual outcomes: Was the name pronounced acceptably? Was the code transcribed exactly? Did the voice remain consistent after the language switch?
Accent consistency also needs an explicit target. My view is that developers should define it as preserving the intended voice and regional choices, not forcing every speaker or language toward one accent. Without that definition, reviewers can disagree about whether a change is an improvement.
Full-duplex behavior warrants its own scenarios. Have a speaker interrupt an answer, correct a detail mid-sentence, or pause before finishing a request. These are proposed validation steps, not results from testing GPT-Live-1.1, but they follow directly from its documented ability to listen while speaking.
For delegation, use clear task boundaries. An application might instruct the voice model to hand account-specific requests to its backend while handling conversational acknowledgments itself. Then verify both the handoff decision and the information transferred. Better transcription is especially consequential when spoken names or numbers become tool arguments.
Standby Changes the Cost of Connected Silence
Standby is available for GPT-Live-1 and GPT-Live-1.1 WebRTC sessions. OpenAI says it pauses inference during inactivity and resumes when someone speaks. The GPT-Live-1.1 pricing documentation lists these duration-based rates:
| Session state | Price per minute | Cost for 60 minutes in that state |
|---|---|---|
| Active | $0.05 | $3.00 |
| Standby | $0.02 | $1.20 |
The standby rate is 60% below the active rate, but that percentage applies only to time billed as standby. A connected session still has a charge while inference is paused.
Frequently Asked Questions
3 questions
1How Do Developers Select GPT-Live-1.1?
Use gpt-live-1.1 in an existing Live session configuration. The documented supported endpoint is v1/live/sessions, not the Realtime or Responses API.