Positioning and release
GPT-Transcribe is OpenAI’s high-accuracy speech-to-text model for completed audio files and Realtime input transcription. OpenAI announced its release on July 28, 2026. (developers.openai.com)
Capabilities and uses
- Transcribes recorded speech in its original language, returning a final transcript or streaming transcript events while a completed file is processed.
- Supports final transcripts of committed turns in Realtime sessions over WebSocket. For low-latency live transcription, OpenAI also offers the distinct
gpt-live-transcribemodel. - Accepts unstructured context, keyword hints, and multiple expected language hints to help with domain terminology, multilingual audio, and code-switching. Detected languages can be returned with the transcript. (developers.openai.com)
- Suitable for transcription of recordings such as meetings or customer-support calls, including recordings containing specialized terms or multiple languages. (developers.openai.com)
Access and limitations
- The file-transcription endpoint is
v1/audio/transcriptions; Realtime transcription uses a Realtime session. File uploads are limited to 25 MB. Supported file formats include FLAC, MP3, MP4, MPEG, MPGA, M4A, OGG, WAV, and WebM. Larger files need compression or splitting. (developers.openai.com) - Audio is input-only; text can be input and output. Images and video are not supported. Streaming is supported, but function calling, structured outputs, and fine-tuning are not. (developers.openai.com)
- Keyword hints are not guaranteed output and can cause unspoken terms to appear. Use
languagesrather than the singularlanguagefield; do not send both. Certain characters in keywords or an overlong prompt cause the request to be rejected. (developers.openai.com) - The file-transcription guide documents speaker labels and word timestamps using specialized legacy models rather than attributing those features to
gpt-transcribe. (developers.openai.com)
Pricing
The OpenAI API model page lists USD $0.0045 per minute of transcription audio. The pricing table labels this as an estimated cost and does not list separate input-token or output-token rates for gpt-transcribe; those separate rates are therefore not verified. Regional-processing endpoints have a 10% uplift where the published regional-processing conditions apply. Pricing checked November 16, 2026. (developers.openai.com)