OpenAI splits speech-to-text into file and live models with GPT Transcribe GA

Developers adding speech-to-text now have to start with a sharper product question: is the audio already bounded, or is it arriving live? OpenAI’s July 28, 2026 release notes made GPT Transcribe and GPT Live Transcribe generally available, splitting file-oriented final transcripts from low-latency transcript deltas for live audio.

The two models carry different cost and latency assumptions. OpenAI’s model catalog lists GPT Transcribe as the high-accuracy speech-to-text model for completed audio files, streamed file transcripts, and committed Realtime turns, priced at $0.0045 per minute. GPT Live Transcribe is the streaming model for applications that need low-latency transcript deltas from live audio, priced at $0.017 per minute. That distinction matters for call centers, meeting tools, voice agents, medical documentation, and media workflows where latency and review quality are not the same requirement.

The notable API change is context. Both models accept free-form prompt context, literal keywords, and a languages list for cases where input may contain more than one language. That is aimed at domain terms, acronyms, multilingual conversations, and code-switching, which are exactly where generic transcripts tend to fail in production.

OpenAI’s transcription guide now recommends gpt-transcribe as the starting model for file transcription and gpt-live-transcribe for realtime transcription. It still keeps specialized paths: speaker-labeled transcripts use gpt-4o-transcribe-diarize, subtitle timestamps and completed-recording translation to English stay with whisper-1, and existing integrations can keep older transcription models where supported. The change is not a full replacement of every audio path, but it resets the recommended default for new builds.

The next test is workflow-specific accuracy rather than broad word error rate. OpenAI’s docs explicitly call out target languages, accents, code-switching, background noise, names, numbers, dates, and alphanumeric strings as test material. As voice becomes an input layer for agents, transcription quality increasingly decides whether downstream tool calls receive clean instructions or expensive ambiguity.