gpt-live-transcribe: realtime STT at $0.017 a minute

gpt-live-transcribe is OpenAI's streaming speech-to-text model, priced at $0.017 per minute, which is $1.02 per hour. It works only on the v1/realtime/transcription_sessions endpoint. If your audio is a finished recording, OpenAI's own guide points you to gpt-transcribe at $0.0045 a minute instead, which is about 3.8 times cheaper.

What the model page says

Facts from OpenAI's model page and transcription guide, read 2026-10-05.

How $0.017 compares with other live options

Streaming prices vary by vendor and by billing basis. AssemblyAI's page notes that streaming is billed on session duration, not audio duration, so a session left open costs money even in silence. The OpenAI listing says per minute; I could not tell from the page whether it uses the same basis, so check before you hold a session open for hours.

When the premium is worth thinking about

On listed numbers, gpt-live-transcribe is the most expensive transcription row above (the translate row is a different product): $1.02 an hour against $0.39 for Scribe v2 Realtime, which is 2.6 times as much (1.02 / 0.39). The price does not say what you get for the difference. The model page describes tunable latency and keyword and language hints, which matter for domain terms and mixed-language speech, but I did not read accuracy or latency numbers for any of these services, so run your own audio through two of them before you pick.

• Use gpt-live-transcribe when you are already inside OpenAI's Realtime stack and want hints for product names.
• Use a file endpoint, such as gpt-transcribe, for anything you can wait to finish.
• OpenAI's guide says speaker labels need gpt-4o-transcribe-diarize, and timestamps or subtitle formats need whisper-1, so a live transcript alone may not give you either.

What Sume offers

Sume does not transcribe live streams. Its speech-to-text is a flag on Video inspect: transcribe: true on one hosted clip, at $0.01 per audio minute, with an optional language_code hint and sentence segments. For a meeting you record and upload afterwards, that works; for captions on a live call, use a streaming vendor.