Overview
GPT-Live-1.1 is OpenAI's full-duplex voice model, reachable in the OpenAI API as the model id `gpt-live-1.1` from November 19, 2026. It takes audio or text as input and returns audio or text as output, and it is served only through the Live endpoint (`v1/live/sessions`). OpenAI's announcement says the model has powered voice in ChatGPT Plus and Pro since October 15, 2026, before it reached the API.
It succeeds GPT-Live-1 in the same line. OpenAI's changelog describes the update as improved pronunciation, accent consistency and transcription in multilingual voice conversations, and the API announcement adds more reliable handling of letters and numbers, better instruction following and more reliable task completion. Like its predecessor it keeps listening while it speaks, and it hands complex reasoning or tool use to a backend model of the developer's choice; developers describe in plain language when that hand-off should happen.
Voice sessions are billed by duration: $0.05 per minute while active and $0.02 per minute in standby, a configuration introduced alongside it for both GPT-Live-1 and GPT-Live-1.1. Standby pauses inference after an idle timeout of 5 to 60 seconds while the WebRTC call stays open, and resumes when someone speaks. Backend model and tool usage is billed separately, and GPT-Live-1.1 has its own concurrent-session limits, separate from GPT-Live-1's.
| Released | 2026-11-19 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Modalities | Audio, Text |
| Status | Generally available |
Strengths
- Full-duplex conversation — listens while it speaks, so callers can interrupt
- Steadier, more natural accents and better pronunciation across languages than GPT-Live-1, per OpenAI
- More reliable handling of spoken letters and numbers, better instruction following and task completion
- Standby billing at $0.02 per minute for idle WebRTC sessions, versus $0.05 per minute active
- Streaming and function calling, with reasoning and actions delegated to a paired backend model
Best for
- Reach for it for multilingual voice agents where a consistent accent and voice across languages matters.
- Reach for it for phone and support flows that read back spelled names, codes and numbers.
- Reach for it for long-open voice sessions with quiet stretches, where standby keeps idle time cheap.
How to access
| Provider | Model ID |
|---|---|
| OpenAI API ↗ | gpt-live-1.1 |
GPT-Live (Voice) — every version
The full lineage of the GPT-Live (Voice) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| GPT-Live-1.1current | 2026-11-19 | — | Proprietary |
| GPT-Live-1 | 2026-09-10 | — | Proprietary |
FAQ
How much does GPT-Live-1.1 cost?
GPT-Live-1.1 voice sessions are billed by duration at $0.05 per minute while active. In standby, a paused WebRTC session costs $0.02 per minute. Backend model usage and tool calls made during the session are billed separately at their normal API rates.
What changed from GPT-Live-1 to GPT-Live-1.1?
OpenAI says GPT-Live-1.1 improves pronunciation, accent consistency and transcription in multilingual conversations, handles letters and numbers more reliably, follows instructions better and completes tasks more reliably. Both models listen while they speak, and they have separate concurrent-session limits in the API.
What is standby in GPT-Live?
Standby is a GPT-Live configuration for WebRTC sessions, available for both GPT-Live-1 and GPT-Live-1.1. After an idle timeout of 5 to 60 seconds of silence, the model pauses inference while the call stays open, and it resumes when someone speaks. Standby time is billed at $0.02 per minute.
Which API endpoint serves GPT-Live-1.1?
GPT-Live-1.1 is served only through the Live endpoint, v1/live/sessions. OpenAI's model card marks Chat Completions, Responses, Realtime, Batch and the other endpoints as unsupported, so existing GPT-Live-1 apps switch by setting gpt-live-1.1 in their Live session configuration.
