Whisper and Audio Models

0:00 For speech to text in the OpenAI API, the current model is GPT-Transcribe (gpt-transcribe), released on Jul 28, 2026 and estimated at $0.0045 a minute, cheaper than the $0.006 a minute of whisper-1, which OpenAI will switch off on Feb 26, 2027. For live captions there is GPT-Live-Transcribe at about $0.017 a minute. The text to speech models (tts-1, tts-1-hd, gpt-4o-mini-tts) shut down on Jan 6, 2027, with gpt-realtime-2.1-mini named as their replacement. Whisper itself also lives on outside the API as open-weight models anyone can download from Hugging Face. For spoken conversations, see Realtime API.

Quick facts

Which OpenAI transcription model should you use?

Sources: OpenAI API pricing, models, changelog and deprecations, read Nov 6, 2026. Per-minute prices are OpenAI's own estimates.

1:27 For new code, use gpt-transcribe for recordings and gpt-live-transcribe for live audio: those are the two replacements OpenAI names for every model being retired (deprecations). The cheapest model today, gpt-4o-mini-transcribe at $0.003 a minute, ends on Feb 26, 2027, so it suits only short-lived work.

GPT-Transcribe

0:25 GPT-Transcribe is OpenAI's "high-accuracy speech-to-text model for file and Realtime input transcription" (models). It runs on two endpoints, v1/audio/transcriptions for files and v1/realtime/transcription_sessions for live turns, and accepts language hints, context and keyword hints to get names and jargon right (model page). In OpenAI's launch video, GPT-Transcribe and GPT-Live-Transcribe both support 57 languages and take a list of words, proper nouns or code terms that matter most. The model page gives no word error rate, so there is no OpenAI figure for how much more accurate it is than Whisper; no independent measurement was found for this page either.

Whisper: the open-source model

Whisper is the speech recognizer OpenAI described in the paper "Robust Speech Recognition via Large-Scale Weak Supervision" (Alec Radford and others), first posted to arXiv on Dec 6, 2022. It was trained on 680,000 hours of audio. OpenAI released the model openly before the paper's arXiv posting, in 2022; the exact release day was not confirmed for this page.

OpenAI's Hugging Face organization lists four generations of the large model (Hugging Face, read Nov 6, 2026):

• whisper-large (page updated Feb 29, 2024)
• whisper-large-v2
• whisper-large-v3 (updated Aug 12, 2024)
• whisper-large-v3-turbo (updated Oct 4, 2024)

The open models keep working after the API's whisper-1 shuts down, because they run on your own hardware. That is the main reason to choose Whisper over GPT-Transcribe: no per-minute bill, and audio never leaves your machine. The trade-off is that you supply the GPU and the serving.

Text to speech

OpenAI's three text to speech models all shut down on Jan 6, 2027:

The replacement OpenAI names is gpt-realtime-2.1-mini (deprecations), a speech-to-speech model in the Realtime API that costs $0.60 per 1M text input tokens and $20.00 per 1M audio output tokens. OpenAI's models page lists no dedicated text to speech model as of Nov 6, 2026 (models); spoken output comes from the realtime models or GPT-Live, which gained prompt-to-voice (a voice described in words) on Oct 16, 2026.

Moving off Whisper and the old models

1. List every call to whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe or gpt-4o-transcribe-diarize.
2. Change recorded-audio calls to gpt-transcribe and live-caption calls to gpt-live-transcribe (deprecations).
3. Move any /v1/audio/translations calls to v1/realtime/translations before Feb 26, 2027.
4. Replace tts-1, tts-1-hd and gpt-4o-mini-tts with gpt-realtime-2.1-mini before Jan 6, 2027.
5. Test speaker labels before switching: the deprecations page names the same two replacements for the diarization model, and neither model page read for this article describes speaker labels.

History

• 2022: Whisper released as open source; paper posted Dec 6, 2022.
• Mar 20, 2025: gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-mini-tts (changelog).
• May 7, 2026: Realtime Whisper and Realtime Translate.
• Jul 28, 2026: GPT-Transcribe and GPT-Live-Transcribe.
• Aug 26, 2026: OpenAI announced the end of whisper-1 and the GPT-4o transcribe models (deprecations).

In ChatGPT, speech goes through Voice Mode and dictation rather than these API models directly. All prices: OpenAI API Pricing.

Frequently asked questions

Is OpenAI Whisper still available? As of Nov 6, 2026, whisper-1 still works in the OpenAI API at $0.006 a minute but shuts down on Feb 26, 2027. The open Whisper models, such as whisper-large-v3-turbo, stay downloadable from OpenAI's Hugging Face page.

What replaces Whisper in the OpenAI API? OpenAI names GPT-Transcribe for recorded audio and GPT-Live-Transcribe for live audio as the replacements for whisper-1, at about $0.0045 and $0.017 a minute.

How much does OpenAI transcription cost? As of Nov 6, 2026, OpenAI estimates GPT-Transcribe at about $0.0045 a minute, whisper-1 and gpt-4o-transcribe at $0.006, gpt-4o-mini-transcribe at $0.003, and the live models at about $0.017.

Does OpenAI still have a text to speech API? Yes, until Jan 6, 2027, when tts-1, tts-1-hd and gpt-4o-mini-tts shut down. OpenAI names gpt-realtime-2.1-mini as the replacement.

Is GPT-Transcribe more accurate than Whisper? OpenAI calls GPT-Transcribe "high-accuracy" but publishes no word error rate on its model page, and no independent comparison was found, so there is no reliable figure yet.

Sources

• Pricing, OpenAI, read Nov 6, 2026
• Models, OpenAI, read Nov 6, 2026
• gpt-transcribe model page, OpenAI, read Nov 6, 2026
• Changelog, OpenAI, read Nov 6, 2026
• Deprecations, OpenAI, read Nov 6, 2026
• Robust Speech Recognition via Large-Scale Weak Supervision, Alec Radford et al., arXiv, Dec 6, 2022
• OpenAI on Hugging Face, Hugging Face, read Nov 6, 2026
• Introducing gpt-transcribe and gpt-live-transcribe, OpenAI, Jul 28, 2026 (video)