Overview
GPT-Live-Transcribe is a low-latency streaming speech-to-text model for live audio. It returns incremental transcript text as speech arrives and a final transcript when an audio turn is committed. It is suited to live captions and transcription of microphone or call audio. (developers.openai.com)
Capabilities and use
- Accepts audio input and produces text; transcription can be configured with free-form context, keyword hints, multiple expected languages, and a latency setting.
- Use it in a Realtime transcription session over WebSocket or WebRTC. Keywords are hints, not guaranteed output. (developers.openai.com)
Limitations
- It does not generate audio or support image or video input. Function calling, structured outputs, and fine-tuning are not supported. (developers.openai.com)
- It does not return word-level timestamps, speaker labels, or transcription confidence scores. In its Realtime transcription sessions, server and semantic voice activity detection are unsupported; clients must commit audio turns. (developers.openai.com)
- Transcripts can contain errors; OpenAI recommends testing with representative audio and target languages. (developers.openai.com)
Pricing
OpenAI lists USD $0.017 per minute of realtime audio for this model. The general pricing table labels this an estimated cost and does not list separate input or output token rates for the model; separate input and output charges are therefore unverified. Pricing source: OpenAI API pricing. Checked November 16, 2026. (developers.openai.com)
Release and official sources
Released July 28, 2026, according to the OpenAI API changelog. See also the model page and Realtime transcription guide. (developers.openai.com)