GPT Transcribe
OpenAI · USA · 2026
OpenAI's recommended transcription model since July 2026 — a quarter cheaper per minute than the Whisper it replaces, though not everything Whisper does came with it.
GPT Transcribe, released on 28 July 2026 together with GPT Live Transcribe, is the model OpenAI now points developers to first when they need speech turned into text. It works on completed recordings through the audio transcriptions endpoint, and also produces the final transcript of a committed turn inside a Realtime session over WebSocket. Billing is by audio time, not tokens: $0.0045 per minute, which is a quarter less than the $0.006 charged for whisper-1 and gpt-4o-transcribe. What sets it apart from the earlier generation is how much the caller may tell it in advance. It accepts free-form context describing the recording, a list of keyword hints — names, product terms, acronyms it would otherwise mangle — and, importantly, a list of expected input languages rather than a single one. The older models take the singular language field and assume one tongue per file; this model takes `languages` and is built for recordings that switch between them mid-sentence. Inside a Realtime session it also reuses earlier transcribed turns as context on its own. Output can be streamed while the file is still being processed. Three things did not move to the new model, and OpenAI says so plainly in its own guide: word-level timestamps and srt/vtt subtitles remain with whisper-1, speaker-labelled transcripts require gpt-4o-transcribe-diarize, and translating a recording into English is still a whisper-1 job. The older transcription models remain available for existing integrations, but are no longer the recommended starting point.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!