What AI meeting notes really cost

Every AI notetaker pricing page shows you a number per user per month. That number is almost never what you pay per month. On 6 September 2026 I went through thirteen of these pages and wrote down the monthly price, the annual price and the catch in the free tier. Then I put my own product in the same table and worked out what the alternative costs. The alternative is to buy the software once and pay an AI company for tokens directly. The arithmetic is below, with every assumption in the open.

The billing toggle

The big number is per user per month, billed annually. Otter's Pro plan shows $8.33 per user per month with a "save 51%" badge; the month-to-month price is $16.99.¹ Fireflies Pro is $10 a month billed annually and $18 billed monthly.² Circleback's page shows Pro at $15 per user per month, which is the yearly-billing rate;³ the price data behind the page has month-to-month billing at $18.⁴ The pattern is a discount for a year's commitment, presented as the price.

Granola is the exception. Its Business plan is $14 per user per month and that is the only way to pay for it.⁵

"Annual plans are not currently available; billing is monthly only."

Granola help centre, Subscriptions and billing, checked 6 September 2026

Google does its own version. Its Workspace pricing page, served to me in Canada, shows Business Standard at CA$18.40 per user per month; that is the cheapest tier with "Gemini in Meet". The line under it reads "All plans billed monthly". The toggle above it reads "Annual (Save 16% with 1 year commitment)".⁶ Billed monthly, committed for a year. I could not get USD figures out of that page from here, so the table below carries the Canadian ones, flagged as such.

The free tier has a catch

Free tiers are the other place to read slowly. Otter Basic gives 300 transcription minutes a month and stops each conversation at 30 minutes.¹ Fireflies' free plan counts transcription credits. How many you start with depends on whether you signed up on the web, in the Chrome extension or on the phone: 3, 5 or 10 for accounts created on or after 1 September 2026. Turning on auto-join lifts the cap.¹⁵ Granola's free plan has no meeting cap, but you can only see notes from the last 30 days. Older notes are kept, and hidden until you upgrade.⁵

tl;dv's catch is the one to read twice. Recordings and transcripts are unlimited and free, but AI notes cover the whole meeting for your first 10 meetings only, then the first 10 minutes of each. That limit is for the lifetime of the account, not per month.¹⁶ Read's free plan is 5 meeting transcripts a month with a 1-hour cap per meeting.¹⁰ Jamie's is 10 meetings a month with a 30-minute limit.¹¹ Circleback's free plan keeps meetings and recordings for 30 days.³ Krisp's pricing page lists no free plan: a 7-day trial, then you pay.⁹

Four are worth naming for the opposite reason. Fathom's free plan is $0 forever with unlimited recordings and transcriptions.⁷ That is a genuinely unlimited free tier, and I have no quarrel with it. MacWhisper's free tier is €0 forever with no minute cap, and transcription runs on the Mac; bringing your own key for the AI, or a local model through Ollama or LM Studio, is a Pro feature.¹² Anarlog's free tier has unlimited on-device transcription and lets you bring your own key for the AI.¹⁴ Meetily's Community edition is open source and free, with the full local transcription engine.¹³ Those last three do all of their transcription on your machine, which is why they can afford to give it away. Krisp is the half-case: English is transcribed on-device, the fifteen other languages on its servers, and it still charges a seat.¹⁷ On the free tiers the language model, when there is one, is either local or paid for by you directly. The paid tiers add a model the vendor hosts: Anarlog Pro includes cloud transcription and a cloud LLM,¹⁴ and Meetily offers a hosted summary provider next to the bring-your-own-key ones.¹⁸

What the meter is really counting

When Hyprnote (now Anarlog) launched on Hacker News in July 2025, a commenter asked why notetakers need "sota LLMs and subscriptions".¹⁹ A local Whisper plus a small open model, they said, "works just fine". It is a fair question, and the answer is written on the pricing pages.

A cloud notetaker pays twice for every minute you record. Once to transcribe the audio on a server, once to run a language model over the transcript. Both bills grow with your talking, so the vendor meters you in the same units: Otter counts minutes, Fireflies counts credits, Read counts meetings, Jamie counts meetings and caps their length. tl;dv is the clearest tell. Recordings and transcripts are unlimited on its free plan; the AI notes are what gets capped.¹⁶ The transcript is cheap for them. The model calls are not.

Move transcription onto your own computer and the first bill disappears. A whisper model runs on a laptop. Then the only metered thing left is the model, and there is no reason the vendor has to be the one paying for it, or marking it up.

The bring-your-own-key arithmetic

Here is the calculation for MidMeeting. Change any assumption you dislike and rerun it.

Each advisor gets the live transcript, and every time someone finishes a sentence MidMeeting checks whether that advisor is due for a look. An advisor is consulted at most once per cadence window. Relaxed is once every 90 seconds, balanced once every 45, chatty once every 20. The four default advisors ship on balanced. That is 80 consultations per advisor per hour on balanced, and 180 on chatty. I run the arithmetic on balanced, and I say so where it matters.

Each consultation sends the last ten minutes of transcript plus a short prompt. Assume people talk at about 150 words a minute, so the window is about 1,500 words, or about 2,000 tokens. Add a running summary of what came before and the advisor's own instructions, and call it 3,000 input tokens per consultation. Assume under 100 output tokens. That is generous: most of the time the advisor decides there is nothing worth writing and answers with a PASS, a handful of tokens. I assume two advisors on. With two advisors on balanced, the hour is 160 consultations, roughly 480,000 input tokens and 16,000 output tokens. These are my assumptions, not measurements of your meetings.

Now the prices, fetched from the providers' own pages on 6 September 2026, per million tokens, standard tier, no caching:

• Anthropic: Claude Haiku 4.5 at $1 input and $5 output; Claude Sonnet 5 at $2 input and $10 output.²⁰
• OpenAI: gpt-5.4-nano at $0.20 input and $1.25 output; gpt-5.4-mini at $0.75 input and $4.50 output.²¹
• Google: Gemini 3.5 Flash-Lite at $0.30 input and $2.50 output; Gemini 3.5 Flash at $1.50 input and $9.00 output.²²

Multiply by 0.48 million input tokens and 0.016 million output tokens:

• Claude Haiku 4.5: $0.48 plus $0.08, so $0.56 for the hour.
• Claude Sonnet 5: $0.96 plus $0.16, so $1.12.
• gpt-5.4-nano: $0.096 plus $0.02, so about $0.12.
• gpt-5.4-mini: $0.36 plus $0.072, so about $0.43.
• Gemini 3.5 Flash-Lite: $0.144 plus $0.04, so about $0.18.
• Gemini 3.5 Flash: $0.72 plus $0.144, so about $0.86.

Those are for two advisors on the balanced cadence, at the token assumption above. The four defaults double every line: $2.24 an hour on Sonnet 5. Chatty is 2.25 times the balanced figure, so $2.52 an hour on Sonnet 5 with two advisors. Relaxed is half. A quiet meeting, or one advisor instead of two, costs less. The running summary that keeps the older transcript in view is refreshed every two minutes and shared by all advisors; it is a small call, and the slack in the 3,000-token assumption covers it.

All three providers also sell prompt caching, which charges a fraction of the input price for the part of a prompt that repeats between calls. Anthropic lists cache reads at $0.20 per million on Sonnet 5,²⁰ and OpenAI lists cached input at $0.075 on gpt-5.4-mini.²¹ Google lists context caching at $0.15 per million tokens on Gemini 3.5 Flash, plus $1.00 per million tokens per hour of cache storage.²² A rolling window repeats most of itself from one consultation to the next, so caching could cut the input line a good deal. I have not assumed it in the numbers above.

A model running on your own computer through LM Studio or Ollama costs nothing per token, and nothing leaves the machine.

Now compare that to a seat. Otter Pro billed monthly is $16.99;¹ Fireflies Pro billed monthly is $18.² At the Sonnet 5 estimate of $1.12 an hour, one Otter seat-month buys about 15 hours of meetings with two advisors on. On Haiku 4.5, about 30 hours. On gpt-5.4-nano, more than 140. And the meter only runs while you are in a meeting. The month you take off costs nothing, and there is no annual commitment to forget to cancel.

The trade-offs

Cheaper comes with chores. Here is what you take on.

• You manage an API account and a key. Signing up, adding a payment method, generating the key, pasting it in. A subscription notetaker does that for you and hides it in the seat price. It takes ten minutes once, but it is ten minutes a notetaker never asks of you.
• The cost varies. A usage bill moves with how much you talk, how many advisors you turn on, which cadence they run at and which model you pick. A seat price is a number you can put in a budget without thinking; a token bill is not.
• There is no team workspace. One licence covers up to three of your own computers. Nothing syncs, nobody else can search your meetings, and there is no admin who can see what the team recorded. If your organisation wants a shared archive, that is not what this is.
• A subscription bundles hosting, storage and support. When you pay $18 a month, some of it buys the server that keeps your recordings and the person who answers tickets. Backups are theirs to configure, not yours. With MidMeeting the audio, the transcript and the summaries are files in a folder on your disk, and backing them up is your job.
• Your computer has to be in the call. A bot can record a meeting you skipped and a phone app can record a meeting in a room. MidMeeting records what your machine hears, so you have to be there with it. There is no mobile app.
• There is no refund. The three free meetings are the trial, so use them on real calls before you pay.

None of this makes the subscription vendors dishonest. They are pricing a real cost: servers transcribing your audio and models reading it, plus the storage and the people around that. That cost is now optional, and the seat price is a choice rather than a fact.

Sources

1. Pricing, Otter.aichecked 6 September 2026
2. Learn about the Fireflies pricing plans, Fireflies.ai help centrechecked 6 September 2026
3. Pricing, Circlebackchecked 6 September 2026
4. pricing.json, the price data behind Circleback's pricing pagechecked 6 September 2026
5. Subscriptions and billing, Granola help centrechecked 6 September 2026
6. Google Workspace pricing, as served in Canadachecked 6 September 2026
7. Pricing, Fathomchecked 6 September 2026
8. Pricing (structured data, EUR per seat), tl;dvchecked 6 September 2026
9. Pricing, Krispchecked 6 September 2026
10. Pricing, Read AIchecked 6 September 2026
11. Pricing, Jamiechecked 6 September 2026
12. MacWhisper, Good Snoozechecked 6 September 2026
13. Pricing, Meetilychecked 6 September 2026
14. Pricing, Anarlogchecked 6 September 2026
15. Learn about transcription credits, Fireflies.ai help centrechecked 6 September 2026
16. What are the limits of the free plan, tl;dv help centrechecked 6 September 2026
17. AI Meeting Assistant, Krispchecked 6 September 2026
18. Summary providers, Meetily docschecked 6 September 2026
19. Launch HN: Hyprnote, Hacker News thread via the Algolia APIchecked 6 September 2026
20. Pricing, Claude by Anthropicchecked 6 September 2026
21. API pricing, OpenAI developer docschecked 6 September 2026
22. Gemini API pricing, Google AI for Developerschecked 6 September 2026