The OpenAI Whisper API cost depends on how much audio you send. The bill grows in a straight line with recording time: an hour of audio costs 60 times the per-minute rate, and 10 hours costs 600 times it. The rate differs by model and changes over time, so we don't quote it here. Check the API pricing page on openai.com before you budget. If you'd rather not pay per minute at all, the open-source Whisper model can run on your own computer for free.
How OpenAI bills transcription
You pay through an OpenAI API key. Every transcription request made with that key is billed to the OpenAI account the key belongs to. The app you use (a script, a dictation tool, a notes app) doesn't set the price. OpenAI does.
The API has more than one transcription model. The ones you'll see in apps are whisper-1 (the original Whisper model) and the newer GPT-4o transcription models. Gubb's settings, for example, list them as "Whisper-1", "GPT-4o Transcribe" and "GPT-4o Mini Transcribe". Prices differ by model, so the model your app sends decides what you pay.
Because the cost follows audio length:
- Pauses still count. A 10-minute recording with 4 minutes of silence is still 10 minutes of audio. Trimming dead air before you upload lowers the bill.
- Re-running a file costs again. If you transcribe the same recording twice to compare models, you pay twice.
Whisper API cost examples by monthly habit
Most people don't know how many minutes of audio they produce in a month. The table converts common habits into monthly minutes. Multiply the minutes by the current per-minute rate for your model.
| How you use it | Audio per month | Monthly cost |
|---|---|---|
| Quick voice notes, 5 minutes a workday (22 workdays) | 110 minutes | 110 × rate |
| Dictating emails and docs, 30 minutes a workday | 660 minutes (11 hours) | 660 × rate |
| Transcribing one 1-hour meeting a week | about 260 minutes (52 weeks ÷ 12 months × 60) | 260 × rate |
| Transcribing a 1-hour recording every workday | 1,320 minutes (22 hours) | 1,320 × rate |
| Dictating 2 hours a workday | 2,640 minutes (44 hours) | 2,640 × rate |
For single recordings, the multipliers are just as simple:
| Audio length | Cost |
|---|---|
| 1 minute | 1 × rate |
| 1 hour | 60 × rate |
| 10 hours | 600 × rate |
| 100 hours | 6,000 × rate |
Work out your own number
- Estimate how many minutes of audio you record on a typical day.
- Multiply by the number of days you record in a month.
- Open OpenAI's pricing page and find the rate for the model your app uses.
- Multiply your monthly minutes by that rate.
- Add some margin for re-runs and recordings you throw away.
Compare it with a dictation subscription
If you're choosing between paying OpenAI directly and paying for a dictation app, find the break-even point. Divide the app's monthly fee by the per-minute rate. The result is how many minutes a month you'd need to record before the subscription works out cheaper.
As of November 2026, third-party pricing pages list Wispr Flow Pro at $15 a month (or $144 a year), with a free plan capped at 2,000 dictated words a week on Mac and Windows. They list Superwhisper Pro at $8.49 a month, with a $249.99 lifetime option. Both apps type into any app on your computer, which a raw API key doesn't do on its own, so the subscription buys more than transcription. See Wispr Flow pricing and Superwhisper pricing for the details.
Local Whisper: no per-minute fee
OpenAI released Whisper as an open-source model, which is why many apps can run it entirely on your computer. Nothing goes to OpenAI, you don't need an API key, and every minute you transcribe costs nothing extra. For background on the model itself, see what OpenAI Whisper is.
Local transcription still has costs, just not on a bill:
- Disk space. Whisper model files are large. In Gubb, "Whisper Small" is 480 MB, "Whisper Medium" is 1,500 MB and "Whisper Large" is 3,000 MB.
- Memory. Bigger models need more RAM. Gubb's settings recommend 8 GB or more for Medium and 16 GB or more for Large.
- Time and processor load. Your computer does the work, so speed depends on your hardware. Superwhisper notes (as of November 2026) that its offline models run really well only on Apple silicon Macs, and that Intel Macs work best with its cloud models.
- A one-time download. You need internet once to fetch the model. After that, transcription works offline.
Ways to run Whisper locally on a Mac without writing code include MacWhisper (its free tier transcribes files with local models, as of November 2026), Superwhisper's free plan (small local Whisper models, according to one review) and Gubb. Apple Dictation isn't Whisper, but on Apple silicon Macs it processes supported languages on the device and costs nothing. The step-by-step version is in how to use Whisper on Mac without coding.
How Gubb handles this
We make Gubb, a notes app with built-in speech to text. Its default engine, Local Models, runs Whisper on your Mac. You download Whisper Small, Medium or Large once under Settings ▸ Voice to Text ▸ Available Models, and transcription then has no per-minute cost. The gubb.net version is a one-time $59 license (as of November 2026), with no AI fee on top.
Know the limits before you rely on it:
- Local transcription in Gubb is English only.
- Only the first 30 seconds of each recording are transcribed. That suits short voice notes, not meetings or long dictation.
- Text goes into the open Gubb note, not into other apps.
- After downloading a model, select it under Active Model (or restart Gubb). Otherwise transcription fails.
Gubb's settings also have an OpenAI API provider with a field for your key. In the current version, that option can't be fully switched on from the settings, so you get "OpenAI API is not configured. Please add your API key in settings." even after saving a key. Use Local Models for now. If you do enter a key, Gubb stores it in plain text in the app's local storage, not in the macOS Keychain.
If short English voice notes are what you need, the Gubb trial download is free for 7 days.
When the API is worth paying for
Pay for the API if:
- Your computer struggles with local models. On an Intel Mac or a machine with little RAM, the API moves the work to OpenAI's servers.
- You transcribe rarely. A few recordings a month cost little, and you skip multi-gigabyte model downloads.
- You need long recordings or other languages and your local tool can't handle them. Check your local tool's limits first. Gubb's local mode, for example, is English only and stops at 30 seconds.
- You're building something. A script or service that needs transcription on a server is simpler to build on an API than on a model you host yourself.
Stay local if:
- You record a lot. Per-minute billing adds up with daily dictation or regular meetings. A local model has no per-minute charge whether you transcribe 10 minutes or 100 hours.
- The audio is private. With the API, your recording is sent to OpenAI. A local model keeps it on your computer.
- You work offline. The API needs an internet connection for every request.
- You want a predictable bill. Local transcription has no usage charges to track.
You can also mix the two: a local model for everyday notes, and the API only for the occasional long or difficult recording. If you're sizing up daily dictation, our guide to the best dictation apps for Mac compares the tools built for that job.
FAQ
Is the Whisper model itself free?
Yes. The open-source model runs on your own computer at no charge. You pay only when you send audio to OpenAI's hosted API.
Do I pay for silence in a recording?
Pauses are part of the audio length, so yes. Trim long silences before uploading to keep costs down.
Does an app add its own fee when I use my OpenAI key?
That depends on the app. With a bring-your-own-key setup, OpenAI bills your account directly for transcription. Some apps also charge for their own features, so check both.