OpenAI ships GPT Transcribe + Live Transcribe, from $0.0045/min
Whisper-1's successor lands, plus a low-latency streaming sibling for live captions — both read keyword hints, language lists, and noisy real-world audio.
Copy markdown
Two models, two jobs
gpt-transcribe handles recorded files and batch jobs; gpt-live-transcribe streams low-latency captions for live speech and voice agents. Both are audio-in, text-out.
The price: $0.0045/min for files
gpt-transcribe runs $0.0045 per minute of audio — under the $0.006/min whisper-1 it replaces — while the real-time gpt-live-transcribe is $0.017/min.
Smarter than whisper on real audio
Both take free-form context, keyword hints, and multiple expected languages, so accents, numbers, specialized terms, and loud background noise land right. OpenAI reports a lower error rate than whisper-1.
Migration is basically one line
Both hit the same v1/audio/transcriptions endpoint — swap model: 'whisper-1' for model: 'gpt-transcribe'. The live model streams over the Realtime API (v1/realtime, transcription_sessions).
Build this weekend
Wire up real-time meeting captions, a voice-agent front end, or a podcast indexer that finally spells names and acronyms right. gpt-live-transcribe beats May's gpt-realtime-whisper on streaming.