A local model against a per-minute API: where the line actually falls
By Andrey ChmerevI sell the local option, so the honest thing is to show the range where it loses — and the second section does. Prices are from OpenAI's own pricing page, read on 5 September 2026. Speed and disk figures are my own measurements on an M4 Max, with the method linked.
Two ways to turn a lot of audio into text: call a speech to text API and pay by the minute, or run local speech to text on your own machine and pay once. Which is cheaper has an actual answer, and it depends almost entirely on how many hours you have.
The rates
From OpenAI’s pricing page, read 5 September 2026:
| Model | Per minute |
|---|---|
| gpt-4o-mini-transcribe | $0.003 |
| gpt-transcribe | $0.0045 |
| Whisper | $0.006 |
| gpt-4o-transcribe | $0.006 |
| gpt-4o-transcribe-diarize | $0.006 |
| gpt-live-transcribe | $0.017 |
Cheap per minute. That is the whole appeal of a hosted Whisper API, and for a lot of people it is the right answer. Other speech recognition API providers price in the same range, so the arithmetic below travels.
Where the API wins, plainly
Against a one-time app at $29, low volume looks like this:
| Audio | at $0.003 | at $0.006 |
|---|---|---|
| 1 hour | $0.18 | $0.36 |
| 5 hours | $0.90 | $1.80 |
| 10 hours | $1.80 | $3.60 |
| 40 hours | $7.20 | $14.40 |
Transcribe ten hours a year and an API costs you between two and four dollars. Buying software for that is a bad trade — and I sell the software.
The break-even lands here:
| Rate | Break-even against $29 |
|---|---|
| $0.003 / min | 161 hours |
| $0.0045 / min | 107 hours |
| $0.006 / min | 81 hours |
| $0.017 / min | 28 hours |
Eighty hours of audio is a lot for one person and nothing for a team that records every meeting. Work out which side of that line you are on, because the rest of this article only matters above it.
What the arithmetic leaves out
Three things appear in neither column, and they decide more cases than the money.
The audio leaves. Every file goes to a third party. For plenty of work that is fine. For interviews under an ethics approval, legal material, HR recordings or anything under an NDA, it decides the question outright, and no price makes it acceptable. Local means the recording never moves — a claim you can verify in a minute by pulling the network out.
The meter changes behaviour. A per-minute charge makes you think twice before transcribing something speculative. That is a real cost that never shows up as a number: the recordings you did not process because it did not seem worth the spend.
The local option has its own bill, in something other than dollars. Models run 200–800 MB each and land on disk compiled, so budget about double the listed size. The first run of each takes minutes while CoreML compiles it — 2 minutes 35 seconds for Whisper large-v3-turbo in my measurements — and seconds every time after.
Accuracy is not the differentiator people expect
The APIs run the same family of models you would run yourself. Whisper is Whisper. Published word error rates on LibriSpeech test-clean put a compressed large-v3-turbo at 1.96%, and that is the model in a local app as much as in a hosted endpoint.
Providers can afford bigger models than a laptop will run, so on hard audio — overlapping speech, strong accents, bad rooms — the gap is real. On clean dictation and decent recordings it is small enough that other things decide.
One place the APIs are genuinely ahead: diarization. gpt-4o-transcribe-diarize labels speakers at the same $0.006 a minute. Local models mostly do not do this at all, and if you need to know who said what from one mixed recording, that alone may settle it.
The case that changed recently: agents
An AI agent that needs to transcribe something used to have one option — call an API, with a key, a bill and an upload.
MCP changes the shape of that. A local MCP server exposes transcription as tools the agent can call, so it works on files on the machine it is already running on. Kekoso ships one with twelve tools: transcribe a file or a link, list and read past transcripts, manage the custom vocabulary, install and remove models.
For an agent the arithmetic above tilts much further, because agents are not careful about volume. One turned loose on a folder of recordings will run up a bill a person would have hesitated over, and a per-minute meter is precisely the wrong pricing model for something that does not feel cost. A local tool has no such failure mode: the worst case is that it takes a while.
And the part that is not about money at all — an agent holding an API key can send your audio somewhere. An agent holding a local tool cannot, whatever it decides to do.
How to decide, in one paragraph
Under about eighty hours of audio and no privacy constraint: use an API and stop thinking about it. Above that, or with material that cannot leave the machine, or with an agent doing the calling: local, and the difference compounds. If you need speaker labels from a single track, the API has a feature the local side does not, and that may outrank everything else.
Prices are OpenAI’s own, read on 5 September 2026; every provider differs. The break-even figures are division you can redo with whatever numbers apply to you.
Questions people ask
How much does a transcription API cost per minute?
From OpenAI's pricing page on 5 September 2026: gpt-4o-mini-transcribe $0.003 a minute, gpt-transcribe $0.0045, Whisper and gpt-4o-transcribe $0.006, and the live variants $0.017. Diarization is $0.006 with gpt-4o-transcribe-diarize, which is worth noting because local models generally do not do it at all.
When is a speech to text API cheaper than running a model locally?
Below roughly 80 hours of audio against Whisper's $0.006 a minute, or 161 hours against the $0.003 tier, comparing with a one-time $29 app. Ten hours a year costs $1.80 to $3.60 through an API. At that volume paying for software is the worse deal, and anyone telling you otherwise is selling something.
Where does the local option win?
Volume, privacy and the absence of a meter. Past the break-even the marginal cost of local transcription is zero, the audio never leaves the machine — which for some material is a requirement rather than a preference — and there is no rate limit, no quota and no bill that scales with a busy month.
Can an AI agent use a local transcription model instead of an API?
Yes, through MCP. A local MCP server exposes transcription as tools the agent can call, so the agent works on files on the machine it is already running on rather than sending them to a cloud endpoint. No API key, no per-minute charge, and the audio stays where it was.
Is a local model as accurate as a transcription API?
For clean speech, close — the APIs run the same family of models. Published word error rates on LibriSpeech test-clean put a compressed Whisper large-v3-turbo at 1.96%, and that is the model in a local app as much as in a hosted endpoint. The gap shows on hard audio, where a provider can afford a larger model than your laptop will run.
What does a local model cost besides the app?
Disk and the first run. Models are 200 MB to 800 MB each and land compiled, so budget roughly double the advertised size. The first compile of a model takes minutes on the machine — 2 minutes 35 seconds for Whisper large-v3-turbo in my measurements — and every run after that is seconds.