—monthly bill
—per audio hour
—is speech-to-text
Where the money goes — per meeting
One meeting, broken into its three billable pieces. Notice how the transcript the model reads and the summary it writes are almost free next to the audio it had to transcribe.
| Component | Size | Cost / meeting | Share |
|---|
Build your own vs a SaaS seat — the break-even
A subscription charges a flat seat fee no matter how much you record; your own pipeline charges per meeting with no floor. Each row is a monthly volume: the API column is what you would pay, the seat column is the flat fee, and the winner flips once you record enough.
| Meetings / month | Your API cost | SaaS seat | Cheaper |
|---|
The summary is free; you are paying for the microphone
Everyone building an internal notetaker braces for a big language-model bill and then discovers the model is the cheapest part of the whole thing. The arithmetic is stark: a 45-minute meeting is only a few thousand spoken words, which is well under ten thousand tokens, and summarising that on a cheap model costs a fraction of a cent — you could run a thousand meetings through the LLM for the price of a sandwich. What actually costs money is the speech-to-text pass, because it is billed per minute of audio and every second has to be processed whether anyone is talking or not. So the lever that matters is not which frontier model writes the summary; it is your transcription rate and the number of meetings you record. That reframes the whole build-versus-buy question. A SaaS seat is a flat fee that makes sense for someone in back-to-back calls all day, but for a team that records a handful of meetings a week, a do-it-yourself pipeline costs a few dollars a month and the break-even is nowhere close. Size the transcription side precisely on the transcription cost calculator, price the summary tokens on the token counter, and compare the flat-fee trade-off directly with the SaaS vs API cost calculator.
How this calculator works
Each meeting is priced in two parts. The speech-to-text cost is the meeting length in minutes times your per-minute transcription rate. The summary cost reads a transcript whose size it estimates from length times words-per-minute times about 1.33 tokens per word, priced at the input rate, plus the summary you asked for priced at the output rate. Adding the two gives cost per meeting; multiplying by meetings per month gives the monthly bill, and dividing by meeting length gives the per-audio-hour figure. The speech-to-text share is that first part over the total. The break-even table re-runs the monthly cost across a range of volumes and compares it to the flat SaaS seat you entered, so you can see exactly where buying beats building.
Frequently asked questions
What does an AI meeting notetaker actually cost to run yourself?
Far less than most people expect — usually a few cents per meeting. A notetaker is two API calls: a speech-to-text pass that turns audio into a transcript, and a language-model pass that turns the transcript into a summary with action items. Speech-to-text is billed by audio minute, so a 45-minute meeting at around a cent a minute costs roughly a quarter to half a dollar. The summary is where people assume the money goes, but a full transcript is only about eight thousand tokens, and summarising it on a cheap model costs a fraction of a cent.
Why is speech-to-text the biggest cost, not the AI summary?
Because transcription is priced by the minute of audio while the summary is priced by the token, and a meeting produces very few tokens relative to how long it runs. People speak at roughly 130 words a minute, so even a long meeting is well under ten thousand tokens for the model to read. The speech-to-text pass has to process every second of audio and is billed accordingly, so it typically accounts for the overwhelming majority of the bill.
Is it cheaper to build my own notetaker or pay for Otter or Fireflies?
It depends on how many meetings you record. A SaaS notetaker charges a flat monthly seat fee whether you record two meetings or two hundred, while a do-it-yourself pipeline charges per meeting with no floor. The API route wins for light and moderate users and the SaaS route wins for heavy ones, and there is a break-even number of meetings where they cross — this calculator prints it directly. Remember the SaaS price also buys integrations, speaker labels and a polished UI.
How do I estimate the token cost of summarising a meeting?
Multiply the minutes by words per minute to get the word count, then by about 1.33 to convert words to tokens — that is what the summary model reads. Add the summary length you want back, in the low hundreds of tokens. Price the input at the model's input rate per million tokens and the output at its output rate. For a typical meeting this is a fraction of a cent, so the money is in the speech-to-text rate and the number of meetings, not the tokens.