OpenAI launches two specialized API models for transcription

Signals Inbox·July 30, 2026·Voice AI

OpenAI has split speech-to-text into two specialist models: one built to transcribe people as they speak, and another built to process completed recordings more accurately and cheaply. The real signal is specialization. Just three months after launching GPT-Realtime-Whisper, OpenAI is already making separate live and batch models the recommended starting points for new transcription apps.

Q1What actually launched?

OpenAI introduced two new transcription models. GPT-Live-Transcribe turns microphone, call, or other live audio into text as speech arrives. GPT-Transcribe handles completed recordings, uploaded files, batch jobs, and committed turns in realtime sessions. Instead of asking one model to balance every trade-off, OpenAI now recommends a different model for each workflow.

Q2Why split transcription into two models?

Because live and recorded audio are different problems. A live model must return words quickly before it knows what the speaker will say next. A file model can use more of the recording before settling on an answer. OpenAI is making that trade-off explicit: GPT-Live-Transcribe prioritizes controllable low latency, while GPT-Transcribe prioritizes accurate final text and cheaper large-scale processing.

Q3How big is the price difference?

GPT-Live-Transcribe costs $0.017 per audio minute. GPT-Transcribe costs $0.0045 per audio minute. That makes the file model roughly 3.8 times cheaper. Transcribing 10,000 hours would cost about $10,200 with the live model and $2,700 with the file model. That gap matters for podcasts, call archives, recorded meetings, media libraries, and any company processing millions of audio minutes.

Q4Is this replacing Whisper?

Not completely, but it changes the default. Existing apps can still use Whisper and OpenAI's earlier GPT-4o transcription models. Whisper also remains useful when developers need word-level timestamps, SRT subtitles, VTT files, or translation of completed recordings into English. But OpenAI now tells developers starting a new transcription integration to begin with GPT-Live-Transcribe or GPT-Transcribe.

Q5What is actually better about them?

Both models accept more useful context. Developers can describe the recording, provide likely keywords such as drug names or product codes, and list expected languages. GPT-Transcribe can also use earlier turns as context during realtime sessions. That targets the errors that often make transcripts unusable even when the general sentence looks correct: names, numbers, acronyms, short phrases, mixed languages, and specialist terms.

Q6Why does this matter now?

OpenAI launched GPT-Realtime-Whisper in May 2026 at the same $0.017-per-minute price now used by GPT-Live-Transcribe. Less than three months later, it is already directing new transcription projects toward a more specialized pair of models. That speed suggests speech-to-text is becoming its own competitive API stack, with separate products for live agents, captions, meeting tools, customer calls, and huge recorded-audio workloads.

Q7So should voice startups care?

Yes. The file model can lower the cost of processing old calls and recordings, while the live model gives developers control over the delay between speech and text. Better keyword and language context could also reduce the embarrassing mistakes that break real products. The bigger move is that transcription is no longer being treated as a single background feature. OpenAI is turning it into a specialized model layer for voice software.