gpt-transcribe: OpenAI's New Speech-to-Text Model
What it does, what it costs, what it leaves out
gpt-transcribe is the model OpenAI now recommends for transcribing completed audio files and batch workloads. It launched on July 29, 2026 alongside gpt-live-transcribe, its low-latency streaming counterpart, and replaces gpt-4o-transcribe as the default choice. This page keeps the verifiable facts in one place — pricing, error rates, context features, and the output types it does not produce.
per minute of audio — 25% below whisper-1
word error rate on Common Voice, 22 languages (whisper-1: 40.37%)
real-time processing speed on completed files
released with gpt-live-transcribe ($0.017/min)
Overview
What gpt-transcribe is
gpt-transcribe is a speech-to-text model served through OpenAI's transcription endpoint (v1/audio/transcriptions) and, for committed turns, the Realtime transcription sessions endpoint. OpenAI's documentation recommends it for file transcription, marks the March 2025 models gpt-4o-transcribe and gpt-4o-mini-transcribe as not recommended for new integrations, and keeps whisper-1 available for specialized capabilities.
The launch announcement emphasises real-world audio: better handling of accents, short phrases, numbers, specialized terminology, and speech recorded over loud background noise. Unlike earlier OpenAI transcription models, it accepts several kinds of context at once and reports which language it detected in the audio.
It returns plain text. Subtitle files, timestamps and speaker labels are not part of its output — that split between raw model and finished transcript is exactly where tools built on top of it earn their place, and it is covered in detail below.
Pricing
gpt-transcribe pricing vs OpenAI's other transcription models
Prices are OpenAI's published API rates per minute of audio, checked July 30, 2026.
| Model | Per minute | Notes |
|---|---|---|
| gpt-transcribe | $0.0045 | Current default for files and batch work |
| gpt-live-transcribe | $0.017 | Streaming / Realtime sessions, tunable latency |
| whisper-1 | $0.006 | Legacy — still the source of timestamps, SRT/VTT and translation |
| gpt-4o-transcribe | $0.006 | Predecessor, not recommended for new integrations |
Accuracy
How accurate is gpt-transcribe?
Two independent reference points: OpenAI's published Common Voice comparison across 22 languages, and Artificial Analysis' word-error-rate benchmark as reported by The Decoder. Lower is better.
| Model | Word error rate (AA benchmark) | Notes |
|---|---|---|
| ElevenLabs Scribe v2 | 2.3% | Benchmark leader |
| Google Gemini 3 Pro | 2.9% | |
| Mistral Voxtral Small | 3.0% | Cheapest of the group at $0.003/min |
| gpt-transcribe | 3.31% | 0.7 points better than its predecessor |
| gpt-4o-transcribe | 4.01% | March 2025 predecessor |
On OpenAI's own Common Voice comparison across 22 languages, gpt-transcribe cuts word error rate to 19.27% against 40.37% for whisper-1 — roughly half the errors on hard multilingual audio. Neither figure comes from testing we ran; sources are linked at the bottom of this page.
Limitations
What gpt-transcribe does not return
These are absences in the API's output, per OpenAI's transcription guide — and the reason a transcript from the raw endpoint is rarely the finished artifact.
No SRT or VTT subtitles
Subtitle response formats remain a whisper-1 capability. gpt-transcribe returns text, not caption files with timing blocks.
No word-level timestamps
Among OpenAI-hosted models, word and segment timestamps come from whisper-1 via timestamp_granularities[] — not from gpt-transcribe.
No speaker diarization
Speaker-labelled transcripts require the separate gpt-4o-transcribe-diarize model; gpt-transcribe does not tell you who spoke.
No translation endpoint
Translating audio into English still runs through whisper-1 on v1/audio/translations.
Context features
Steering it: prompt, keywords, languages
gpt-transcribe accepts three kinds of optional context. Hints are suggestions rather than instructions — a keyword only appears in the transcript if the audio actually contains it.
prompt
Free-form text about the recording — who is speaking, what it covers — that helps the model resolve ambiguous audio.
keywords
Literal terms likely to occur: product names, acronyms, people. Useful precisely where generic models guess wrong.
languages
A list of expected input languages (plural — older models accept a single language hint). The response reports what was detected, which helps with code-switching audio.
Minimal API call
The model is API-only; there is no OpenAI-hosted web page for running it on a file. A minimal request looks like this:
curl https://api.openai.com/v1/audio/transcriptions \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-F model="gpt-transcribe" \
-F file="@meeting.mp3"No-code route
Use it online, without an API key
If what you actually need is a finished transcript — with subtitles, speaker labels and timestamps — the raw endpoint leaves that assembly to you. GPT Transcribe, the tool this site is built around, is the browser route: upload audio or video, record live, or paste a media link, and export the result as TXT, SRT, VTT, DOCX, JSON or PDF.
Transcription on this site deliberately runs OpenAI's open-source Whisper model with a separate diarization stage rather than the gpt-transcribe endpoint, because that pipeline is what produces the timestamps, subtitle files and speaker labels the endpoint does not return. New accounts start with five free minutes.
GPT Transcribe (gpt-transcribe.org) is an independent product and is not affiliated with, sponsored by, or endorsed by OpenAI. gpt-transcribe is OpenAI's model; this page documents it.
FAQ
gpt-transcribe, answered briefly
What is gpt-transcribe?
OpenAI's speech-to-text model for completed audio files, streamed file transcripts and committed Realtime turns, released July 29, 2026. It is the recommended default for transcription, succeeding gpt-4o-transcribe.
How much does gpt-transcribe cost?
$0.0045 per minute of audio through the OpenAI API — 25% below the $0.006 rate of whisper-1 and gpt-4o-transcribe. The streaming model gpt-live-transcribe costs $0.017 per minute of session audio.
Is gpt-transcribe better than Whisper?
On accuracy, clearly: OpenAI's Common Voice comparison across 22 languages shows 19.27% word error rate against whisper-1's 40.37%. On capabilities, not entirely — whisper-1 still provides word timestamps, SRT and VTT output, and the translation endpoint, none of which gpt-transcribe offers.
What is the difference between gpt-transcribe and gpt-live-transcribe?
gpt-transcribe is optimised for asynchronous transcription of completed files and batch workloads; gpt-live-transcribe is built for low-latency live transcription over the Realtime transcription endpoint, with tunable latency. They launched together on July 29, 2026.
Does gpt-transcribe support speaker labels or subtitles?
No. Speaker diarization requires the separate gpt-4o-transcribe-diarize model, and subtitle formats with timestamps come from whisper-1. That gap is what this site's own tool closes by running Whisper with a diarization stage.
Can I try gpt-transcribe without writing code?
OpenAI ships it as an API only, so using the model itself means an account, a key and integration code. For a no-code transcript with subtitles and speaker labels included, this site's browser workspace does the whole job with five free minutes to start.
Sources
Where these facts come from
Figures on this page were checked against the following sources on July 30, 2026.