OpenAI cuts batch speech-to-text to $4.50/1k min, adds a live model
GPT Transcribe undercuts GPT-4o by 25% at 3.31% error, while GPT Live Transcribe streams captions at $0.017/min with tunable latency and keyword hints.
Copy markdown
Two models, two jobs
GPT Transcribe handles completed files, streamed file transcripts, and committed Realtime turns over WebSocket; GPT Live Transcribe emits low-latency deltas from live audio. Batch for accuracy, live for captions and voice agents.
Batch STT drops to $4.50 per 1,000 minutes
GPT Transcribe runs $0.0045/min — 25% below GPT-4o Transcribe — putting OpenAI level with Deepgram and AssemblyAI on price. It scores 3.31% on Artificial Analysis's word-error benchmark (#9, 0.7 points better than its predecessor).
Live streaming lands at $0.017/min
GPT Live Transcribe runs through the Realtime API at $0.017 per minute of audio, with tunable latency so you can trade a little speed for accuracy. That's roughly 4x the batch rate — reach for it only when you actually need real-time.
Context that actually improves the transcript
Both take three kinds of hints: an unstructured prompt describing the topic or setting, literal keyword hints for product names, jargon and numbers, and multiple language hints for multilingual or code-switching audio. This is the fix for domain terms Whisper mangles.
Tuned for messy, real-world audio
OpenAI says both models handle accents, short phrases, numbers, specialized terminology, and loud background noise better than before — exactly the failure modes that wreck call-center, field-recording and podcast transcripts.
Build this weekend
A meeting-notes tool (batch the recording, prime it with attendee names as keywords), live captions for a stream, or a voice-agent front end piping Live Transcribe deltas straight into your LLM. Same API key you already have — no new SDK.