GPT-Transcribe: OpenAI's New Speech-to-Text Model

What Is GPT-Transcribe

GPT-Transcribe is the speech-to-text model OpenAI released on July 28, 2026, alongside its streaming counterpart gpt-live-transcribe. It transcribes completed audio files, streams partial text while a file is processing, and handles committed turns in Realtime sessions. OpenAI's transcription guide now opens with one instruction: start with gpt-transcribe. Whisper had held that spot since 2022.

Accuracy is where the difference shows. On OpenAI's launch benchmark, Common Voice across 22 languages, word error rate drops from 40.37 percent with whisper-1 to 19.27 percent with GPT-Transcribe. That is less than half the mistakes on identical audio. The hint system targets the terms that dictation gets wrong in practice: domain vocabulary, multilingual audio, and sentences that switch language midway.

The context features run through three request fields on the same /v1/audio/transcriptions endpoint Whisper uses: prompt for a free-form description of the recording, keywords for literal terms you expect to hear, and languages for expected input languages. The response reports which languages the model actually detected. It accepts files up to 25 MB in mp3, mp4, m4a, wav, webm, mpeg, and mpga.

GPT-Transcribe Pricing

File transcription with GPT-Transcribe costs $0.0045 per minute of audio, per the official model page. That is 25 percent below the $0.006 per minute that both whisper-1 and gpt-4o-transcribe charge, so the accuracy upgrade also costs less. An hour of audio runs about 27 cents; three hours of dictated audio in a day stays under a dollar. The realtime model is the expensive one: gpt-live-transcribe bills $0.017 per minute of session audio, about $1 per hour.

There is no free tier for any hosted OpenAI transcription model. For zero cost rather than maximum accuracy, the roundup of free speech-to-text APIs covers what Groq, Deepgram, and others give away, and Whisper run locally costs nothing at all.

GPT-Transcribe vs Whisper and GPT-4o

Against whisper-1, the new model halves the word error rate, costs 25 percent less, streams output for files, takes richer context than Whisper's 224-token prompt cap allows, and accepts multiple language hints instead of a single language field. Whisper keeps two specialist jobs. It is still the only hosted OpenAI model that returns word and segment timestamps for subtitle files, and the only one on the translation endpoint that turns foreign-language audio into English text.

GPT-Transcribe succeeds gpt-4o-transcribe, the March 2025 model, at a lower price and with a bigger context toolkit: keyword lists, several language hints at once, and detected-language output on top of the prompt field the 4o models already supported. The 4o generation remains available for existing integrations, and gpt-4o-transcribe-diarize remains the pick when speaker labels matter. New projects have no reason to start on gpt-4o-transcribe.

One caveat applies to all of them: these are cloud models. Audio leaves your machine and is processed on OpenAI's servers. For anyone that rules out, the Parakeet vs Whisper comparison shows what on-device engines achieve with no network connection at all.

How to Use GPT-Transcribe in Spokenly

Spokenly supports GPT-Transcribe on macOS, iOS, and Windows from day one, through two routes. Either way, the model ends up bound to a system-wide hotkey: press it, speak, and the transcript lands at the cursor of whatever app has focus.

Route 1: built into Spokenly Pro

Select GPT Transcribe in Spokenly's model list, in the Most Accurate category. No OpenAI account, no API key, no per-minute billing; usage is covered by the Pro subscription. Audio on this route is still cloud-processed, by Spokenly's backend and OpenAI; for dictation that never touches a server, Spokenly's local models with Local Only Mode keep everything on-device.

Route 2: bring your own API key

Bring-your-own-key works on Spokenly's free plan with no word caps. Create an API key in your OpenAI dashboard, open Spokenly's model settings, choose the OpenAI provider, and paste the key. Spokenly preselects gpt-transcribe, so nothing else needs configuring. You pay OpenAI directly: $0.0045 per minute of audio with the file model, $0.017 per minute with realtime streaming.

Spokenly layers its own features on top of either route: automatic language detection across 100+ languages, custom word replacements for terms the model still misses, and AI formatting prompts that strip filler words. MCP integration pipes the same dictation into Claude Code, Cursor, and Codex.

Real-Time Dictation with GPT-Live-Transcribe

The second launch model, gpt-live-transcribe, targets live audio. Instead of waiting for a finished recording, it holds a WebSocket session and emits transcript deltas as you speak, with latency you can tune toward speed or stability. OpenAI lists it in the same top performance tier as the file model, at $0.017 per minute of session audio.

In Spokenly, enable streaming on the OpenAI provider and dictation switches to gpt-live-transcribe automatically: words appear on screen while you are mid-sentence rather than after you stop. For long-form dictation, seeing the text as it is spoken lets you catch a misheard term in the moment instead of after the fact. Anyone who configured the interim gpt-realtime-whisper model keeps a working setup, while new configurations select the live model by default.

Picking the Right OpenAI Model

GPT-Transcribe FAQ

Read Next

• Free speech-to-text APIs compared
• Parakeet vs Whisper: local model showdown
• How to run Whisper locally
• Voice dictation for developers
• Dictation for Mac