gpt-realtime-whisper is a streaming transcription model that converts live audio into text as the audio arrives, rather than waiting for a complete utterance or file upload. It is exposed as a Realtime API model that maintains a persistent streaming session, which makes it suited to continuous, low-latency scenarios such as live captioning, quality monitoring, or feeding a speech-to-speech or translation pipeline in parallel. The model is distinct from the offline transcription path served by POST /v1/audio/transcriptions on gpt-transcribe, so it is the right choice when transcript updates are needed while audio is still in progress rather than as a batch job.

Released alongside two other OpenAI Realtime Voice API models in early May 2026, gpt-realtime-whisper slots into a family of audio-first endpoints designed for voice agents and live audio stacks. Its practical fit is narrow but well defined: pair it with companion realtime models when you need a continuous source-language transcript flowing alongside translation, and reach for the offline gpt-transcribe endpoint instead when you only have pre-recorded files. For builders working on real-time voice experiences, this gives a cleaner separation between always-on streaming recognition and one-shot file transcription, with temperature-style controls available to tune output behavior.