Tasks and inputs
GPT-Live-1 is listed for Audio-to-audio, Text → Speech, Speech → Text. It accepts audio, text and produces audio, text. Use the publisher’s input format and preprocessing for your chosen deployment.
GPT-Live-1 is OpenAI's full-duplex voice model on the Live API, talking and listening at once while a backend model handles reasoning, at $0.05 per minute.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Indicative: $0.05 per minute of Live session, billed per second (not rounded up). Backend model and tool usage billed separately. Prompt-generated voices are free to create. Confirm on your OpenAI account.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Phone agents fail in a particular way. The caller interrupts, the bot keeps talking, and two seconds of overlap turn a support call into a shouting match. GPT-Live-1 is OpenAI's attempt to fix the turn-taking itself: a full-duplex voice model that can listen and speak at the same time, while a separate backend model or agent does the thinking and calls the tools.
It went generally available on the API on 10 September 2026, after launching in ChatGPT on 8 July.
This split is the design decision that shapes everything else. The voice layer handles conversation: hearing, speaking, interruptions, pacing. Reasoning and tool use are delegated to whichever backend you configure. The conversation can carry on while that backend works.
It explains an odd detail in the spec. The voice model's own knowledge cutoff is 31 July 2025, much older than OpenAI's current text models. That matters less than it looks, because facts and decisions should come from the backend, not from the voice layer.
| Field | Value |
|---|---|
| API ID | gpt-live-1 (single snapshot) |
| Endpoint | v1/live/sessions only |
| Input / output | Audio and text in, audio and text out |
| Features | Streaming, function calling |
| Not supported | Structured outputs, fine-tuning, predicted outputs |
| Knowledge cutoff | 31 July 2025 |
It does not run on Realtime, Responses or Chat Completions. Some gateways list a 128K context and 4,096-token output for it; OpenAI states neither, so treat those as a gateway's own limits, not the model's.
The price is $0.05 per minute of Live session, billed per second and not rounded up. An hour of talk is three dollars of voice. The backend model and any tools are billed on top at their normal rates, so a call that leans on a frontier reasoning model can cost far more in backend tokens than in voice minutes. Budget the two separately. These rates are indicative; confirm them on your account.
The API has moved quickly since September:
Outbound SIP is the one that turns a demo into a contact-centre product.
Limits are measured in simultaneous sessions, not tokens: 50 at the Build tier, 300 at Launch and 500 at Grow. There is no free-tier access. If you expect a Monday-morning call spike, check which tier you are on before launch day.
OpenAI also sells a speech-to-speech Realtime family, such as GPT-Realtime-2.1. The practical difference is architecture. Live is built around handing reasoning to a backend you pick, so the same voice can front a fast mid-tier model for routine calls or a frontier one such as GPT-6.1 Sol for hard cases. If you would rather one speech-to-speech model handle the whole exchange, test the Realtime family against it.
Record twenty real calls from your support line, anonymised. Replay the caller's side against a Live session with your intended backend. Measure three things: how often it talks over the caller, how long silences last while the backend works, and total cost per call including backend tokens. The second number is usually the one that decides whether customers hang up.
Other speech models are grouped under audio-to-audio.
Compare the model on the work you need to complete.
GPT-Live-1 is listed for Audio-to-audio, Text → Speech, Speech → Text. It accepts audio, text and produces audio, text. Use the publisher’s input format and preprocessing for your chosen deployment.
Compare pronunciation, voice consistency and time to the first playable audio. Check the service’s billing unit for the actual script.
Use the model identifier and account requirements for the selected API route. Confirm its request limits and output format with one complete task before connecting a larger workflow.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
OpenAI
GPT-Live-1.1 is OpenAI's full-duplex Live API voice model, with more accurate transcription and steadier multilingual accents than GPT-Live-1.
Microsoft
MAI Realtime 1 is Microsoft AI's first speech-to-speech model for live voice agents, from $0.01 per audio minute, in limited preview via MAI Labs and Foundry.
rime
Rime Prism is a speech-to-speech model for enterprise voice agents that settles each reply as inspectable text first, with parallel tool calls.
Liquid AI
LFM3-Audio-2B is Liquid AI's open 2B speech model: ASR, TTS, voice chat and interleaved text-and-speech output from one local checkpoint.
Alibaba
Qwen-Audio-3.1-Realtime Flash is Alibaba's lighter full-duplex speech model with audio and text in and out, function calling, web search and a 262K context.
ElevenLabs
Eleven Realtime v1 Max is ElevenLabs' reasoning speech-to-speech agent for support and multi-step tool workflows, OpenAI Realtime-compatible.