What is GPT Live Transcribe?
GPT Live Transcribe (gpt-live-transcribe) is OpenAI's model for transcribing audio while it is still arriving, such as a microphone feed, a phone call or live captions. It was released on 28 July 2026 alongside GPT Transcribe, its counterpart for finished files, and replaces gpt-realtime-whisper as OpenAI's default for live transcription. OpenAI names it as one of the two replacements for whisper-1 and the GPT-4o transcription models, which shut down on 26 February 2027.
It runs only inside Realtime transcription sessions over WebSocket or WebRTC and returns text, never audio. A delay setting with five levels, from minimal to xhigh, trades how soon partial text appears against accuracy. Sessions accept a free-form context prompt, a keyword list and several language hints, and can be updated mid-session.
How much does GPT Live Transcribe cost?
| Plan | Price | What you get |
|---|---|---|
| gpt-live-transcribe | $0.017 per minute | Realtime transcription sessions only |
| gpt-transcribe | $0.0045 a minute | Finished files |
$0.017 per minute of streamed audio (about $1.02 an hour), the same as gpt-realtime-whisper. No Batch. The file model gpt-transcribe costs $0.0045 a minute.
Is GPT Live Transcribe worth it?
GPT Live Transcribe is a modest upgrade on gpt-realtime-whisper at the same price. On OpenAI's own figures, its transcription error rate on real-world audio in nine languages falls from 11.65% to 9.60%, about 18% fewer errors, while on Common Voice across 22 languages the gain is under a point, 20.33% to 19.70%. Adding a context prompt raised OpenAI's meaning-based accuracy score from 38.5% to 44.6%. These are vendor benchmarks; OpenAI publishes no latency figures.
The price is the weak point. At $0.017 a minute, about $1.02 an hour, it costs almost four times as much as GPT Transcribe on finished files at $0.0045, and nearly twice Microsoft's introductory $0.54 an hour for MAI-Transcribe-2-Streaming, which took first place on Artificial Analysis's streaming leaderboard in October with 2.5% WER at 0.13 seconds after the end of speech. A product that can wait a few seconds for text should not pay the live rate.
The model does less than its rivals on its own. It does not support server-side voice activity detection, so your application must decide when a speaker has stopped and commit each turn. It returns no word timestamps, speaker labels, confidence scores or detected language; for those OpenAI points back to whisper-1 or gpt-4o-transcribe-diarize, both of which shut down on 26 February 2027 with no like-for-like replacement announced. Final transcripts can arrive out of order and must be matched by item ID.
The controls are sound. Five delay levels let a captioning app show text early and a note-taker wait for accuracy, keywords help with product names, and a malformed keyword rejects the whole session update, which is strict but easy to catch in testing. Rate limits start at 2,000 requests a minute on the Build tier and reach 10,000 on Launch and Grow.
Data handling suits regulated buyers. API audio is not used for training by default, /v1/realtime keeps abuse-monitoring logs for 30 days and is eligible for zero data retention, and gpt-live-transcribe is supported for data residency in the US and Europe.
Where it earns its place
A front-end engineer adding live captions to a webinar platform
Show captions within a second or two of speech
Open a Realtime transcription session over WebRTC, set delay to low, and run voice activity detection in the client to commit turns. A 60-minute webinar costs about $1.02 per audio stream.
A backend engineer migrating a contact-centre app off gpt-realtime-whisper
Cut transcription errors without changing the pipeline
Keep the Realtime connection and change the model, replacing the single language field with the languages list and adding product names as keywords. The price per minute stays the same.
A product manager at a European telehealth start-up
Transcribe consultations live without audio leaving the EU
Use EU data residency and apply for zero data retention on /v1/realtime. Speaker attribution between clinician and patient is not provided, so separate channels or a later diarisation pass are needed.
Before you commit
- Setup
- Connect to wss://api.openai.com/v1/realtime?intent=transcription, send session.update with type transcription and model gpt-live-transcribe, stream 24 kHz PCM audio and commit each turn yourself.
- Integrations
- Realtime transcription sessions over WebSocket or WebRTC only; not the file transcription endpoint, Chat Completions, Responses or Batch. Also listed on Azure AI Foundry and Vercel AI Gateway.
- Data and privacy
- API audio is not used for training by default. /v1/realtime keeps abuse-monitoring logs for 30 days and is eligible for zero data retention; data residency is available in the US and Europe.
- Cost at scale
- 100,000 live minutes a month cost $1,700 with no volume discount; the same audio as finished files on gpt-transcribe costs $450.
- Limits
- 2,000 requests a minute and 210,000 tokens a minute on Build, 10,000 requests on Launch and Grow. No server VAD, timestamps, speaker labels, confidence scores or language detection; one invalid keyword rejects the session update.
What’s good
- Five delay levels, from minimal to xhigh, to trade speed for accuracy
- Context prompt, keyword hints and several languages per session
- About 18% fewer errors than gpt-realtime-whisper on OpenAI's real-world test, at the same price
- Eligible for zero data retention, with US and Europe data residency
- Drops into existing Realtime sessions with a model change
What’s not
- $0.017 a minute, nearly four times the file model
- No server-side turn detection; you must run your own
- No timestamps, speaker labels, confidence scores or language detection
- No published latency figures
- Only works in Realtime transcription sessions
Best for
Teams already on OpenAI's Realtime API who need live captions or transcripts in several languages and can do without speaker labels.
GPT Live Transcribe alternatives
GPT Live Transcribe: common questions
- Is GPT Live Transcribe worth it?
- A dependable default for live captions on OpenAI's stack, priced like a premium model while leaving speaker labels, timestamps and even turn detection for you to build. We rate it 3.7 out of 5.
- How much does GPT Live Transcribe cost?
- $0.017 per minute of streamed audio (about $1.02 an hour), the same as gpt-realtime-whisper. No Batch. The file model gpt-transcribe costs $0.0045 a minute.
- What are the downsides of GPT Live Transcribe?
- $0.017 a minute, nearly four times the file model; no server-side turn detection; you must run your own; no timestamps, speaker labels, confidence scores or language detection; no published latency figures; only works in Realtime transcription sessions.
- Who is GPT Live Transcribe best for?
- Teams already on OpenAI's Realtime API who need live captions or transcripts in several languages and can do without speaker labels.
- How much does GPT Live Transcribe cost for a team or in production?
- 100,000 live minutes a month cost $1,700 with no volume discount; the same audio as finished files on gpt-transcribe costs $450.
- Is GPT Live Transcribe safe for company data?
- API audio is not used for training by default. /v1/realtime keeps abuse-monitoring logs for 30 days and is eligible for zero data retention; data residency is available in the US and Europe.
- What are the best GPT Live Transcribe alternatives?
- The closest alternatives we have reviewed are MAI-Transcribe-2-Streaming, GPT Transcribe, ElevenLabs Scribe v2 and Deepgram Nova-3.
Answers checked Nov 6, 2026.