What speech to text does ChatGPT use?

OpenAI has not published the speech model behind the ChatGPT dictation button. The release notes describe a new model without naming it, and the developer documentation names models for the API, which is a separate product from the app. Whisper is the model family people usually mean when they ask, and OpenAI does make it, but nothing we read says ChatGPT dictation runs on Whisper. The honest answer is that the model is undisclosed, and the answer depends on whether you mean the app, the API or the voice feature.

The app's dictation model is not named in OpenAI's release notes

OpenAI's release notes record a new speech-to-text model for dictation in ChatGPT, rolled out from June 26, 2026, across all plans, with no change to how users dictate ¹. The note names no model. It gives one number, and that number is a vendor's own test. OpenAI says its evaluations found word error rate "at least 10% lower" for the top languages it tested than with the previous production model ². The note does not name the languages, the test set or the previous model, so outsiders cannot repeat the test.

We therefore cannot say which model dictation uses on a given date. Treat the June change as the one documented model event for the feature, and do not assume the name of any model from its general reputation.

Whisper is OpenAI's speech model, and the API names it for translation

Whisper is a speech recognition model that OpenAI trained on a large body of audio. The original paper describes training on 680,000 hours of multilingual and multitask audio with transcripts ³. That is the origin of the name people search for when they ask about ChatGPT.

OpenAI's speech-to-text guide still names whisper-1, but for one job: "To translate a completed audio recording into English, use /v1/audio/translations with whisper-1" ⁴. That is a developer feature that returns English text from a recording in another language. It is not a statement about the ChatGPT button. Reading it as proof that ChatGPT dictation runs on whisper-1 would be a mistake.

The developer route for transcription uses gpt-transcribe

For transcription in the API, the guide's quickstart sends the audio to the transcriptions endpoint with the model gpt-transcribe ⁵. The guide also documents gpt-4o-transcribe, gpt-4o-mini-transcribe and a diarization model for speaker labels. Prices are listed separately. OpenAI's price list shows gpt-transcribe at $0.0045 per minute and gpt-4o-transcribe at $0.006 per minute ⁶.

If you are building something, these are the names to use, and the guide and price list are the sources to check on the day you build. They describe the API. They do not tell you which model the ChatGPT app uses, so do not carry the API model name over to the app.

Claude's API does not offer audio input at all

For comparison, Claude's models are text and image in, text out. The models overview lists no audio input ⁷. A developer who wants speech to text with Claude needs a separate speech service, and the text can then go to Claude.

Why the question has no single answer

The word "ChatGPT" covers several products, and each one can use a different path:

The honest answer for the app is "not published." If you need the model for a specific reason, such as a compliance review, ask OpenAI directly, and record the date and the answer you were given.

What this means for your dictation

The model name matters less than what you see. Test the button on the text you care about, and check the words before you send. The transcription is a draft, and the errors that matter are the ones in names, numbers and technical terms. For a general view of how speech models make errors, see our guide to speech to text for AI, and for the difference between the app and the voice feature, read the difference between ChatGPT dictation and voice mode.

For the rest of the picture, our guide to dictation in ChatGPT and Claude covers how each button works, and the apps and browsers overview lists the related guides. If the button itself is failing, see why ChatGPT dictation is not working.