OpenAI releases GPT Transcribe and GPT Live Transcribe

OpenAI added two new speech-to-text models to its API on July 28, 2026: GPT Transcribe for file-based transcription and GPT Live Transcribe for low-latency streaming transcription, rounding out a month that already brought new realtime voice and flagship text models to the API lineup.

What's new

Per OpenAI's own changelog, the company "released GPT Transcribe for accurate file transcription and final transcripts of committed Realtime turns, along with GPT Live Transcribe for low-latency streaming transcription." The two models split the transcription workload by use case:

• GPT Transcribe — targets accurate, after-the-fact transcription of uploaded audio files, and also produces the final, cleaned-up transcript once a Realtime API turn is committed.
• GPT Live Transcribe — built for streaming: low-latency, word-by-word transcription while audio is still being captured, aimed at live captioning and voice-agent pipelines that need to react as a user speaks.

Both models ship with the same feature set: free-form transcription context (developers can pass background text to bias the output), keyword hints (for names, jargon, or product terms likely to appear), and support for multiple expected input languages in a single request. They're available through the existing v1/audio/transcriptions endpoint for file uploads and v1/realtime for streaming sessions.

Context

The release caps a busy month for OpenAI's audio and voice stack. On July 6, the company shipped GPT-Realtime-2.1, an updated realtime voice model with better alphanumeric recognition and interruption handling, plus a distilled mini variant for cheaper voice apps. On July 9, the GPT-5.6 family (Sol, Terra, Luna) landed with programmatic tool calling and multi-agent orchestration features. A week later, on July 22, OpenAI added hard organization-level spend caps that return a 429 once a monthly budget is hit. GPT Transcribe and GPT Live Transcribe fill the remaining gap in that stack: dedicated, purpose-built transcription models rather than transcription-as-a-byproduct of a realtime voice call.

Why it matters

Splitting transcription into a batch model and a streaming model is a direct response to how developers actually build with speech: call-center analytics and meeting transcription need accuracy on a finished recording, while live captioning, voice agents, and dictation tools need results as words are spoken, even at some accuracy cost. By shipping both under one keyword-hint and multi-language feature set, OpenAI is competing more directly with dedicated transcription vendors like ElevenLabs and Deepgram rather than treating speech-to-text as a side effect of the Realtime API. For teams already building voice agents on OpenAI's stack, it also removes a common workaround — routing audio to a third-party transcription API alongside the Realtime API — collapsing that into a single vendor.