How much does the OpenAI speech-to-text API cost?

The OpenAI speech-to-text API bills by the minute of audio for its transcription models. Of the batch models in the table, the cheapest is gpt-4o-mini-transcribe at $0.003 per minute ¹, and the highest is gpt-4o-transcribe at $0.006 per minute ². Two details can change what you pay: the per-minute figures are labelled as estimates, and the same price table lists token prices. The speech-to-text API guide puts OpenAI next to the other providers, and the overview of speech recognition covers the subject more broadly.

The batch models cost $0.003 to $0.006 per minute

The table gives the list rate per minute for each batch model.

To estimate a monthly bill, multiply your audio minutes by the rate for the model you choose. If your volume is in hours, convert to minutes first.

The per-minute figures are labelled as estimates

OpenAI's price table puts its per-minute speech rates in a column headed Estimated cost ³. The same rows also show token prices. For gpt-4o-transcribe the table gives $2.50 input and $10.00 output per 1 million tokens, and for gpt-4o-mini-transcribe $1.25 input and $5.00 output ⁴.

The table does not explain how many tokens a minute of audio uses. So the per-minute rate is OpenAI's estimate, and the token prices are the rates you can check against your own usage. Run one of your files through the model and compare the usage on your account with the minute estimate before you commit to a budget.

Live transcription costs $0.017 per minute

OpenAI's price list gives gpt-realtime-whisper, for live transcription, at $0.017 per minute ⁵. It is a streaming model for live audio, such as captions. For uploaded recordings, compare the batch models in the table above.

We found no whisper-1 price on the list

Many people search for the Whisper API price. The price list we read on 2026-10-17 has no whisper-1 row. The speech-to-text guide still names whisper-1 as the model for translating a completed recording into English ⁶, so the model is in use for that endpoint. We could not confirm a transcription price for it on the pages we read.

If your code calls whisper-1, check the current price list or your billing page before you rely on a figure. The open Whisper code and model weights use the MIT License ⁷, so you can run them yourself with no per-minute bill, though you need hardware. The Whisper guide explains the model.

Data terms are separate from the price

OpenAI says that data sent to its API is not used to train or improve its models, unless the customer explicitly opts in ⁸. Those terms sit on the platform's data controls page, not on the price list, so read them with the price if your audio is sensitive.

What to do

1. Pick the model by what you need. Use a batch model for files, and gpt-realtime-whisper only for live audio.
2. Test on your own audio. Send 30 minutes of recordings to two models, then measure word error rate against a reference transcript you wrote yourself.
3. Check the token usage. Compare the bill for one test file with the minute estimate, because the table does not say how audio becomes tokens.
4. Compare the other providers. The free options for speech-to-text APIs and the Google price question use the same per-minute basis, so you can put the rows side by side.