Daniel Welsh · the calculator · Granola

Granola’s stack,
priced by the hour.

I pulled the Granola Mac app apart and priced every service it calls at the vendor’s public list price. One hour-long meeting costs about 51¢ to $1.03 to run, and 89 to 99% of that is turning speech into text. The AI summary costs pennies.

* An estimate from public list prices, not Granola’s real bill. Granola hasn’t published its costs, discounts or user numbers. App version 7.663.2 for Mac; prices checked 18 Oct 2026.

58¢

speech to text, per meeting-hour

2

transcription streams per meeting

$4.60

per user a month, at list prices

3 to 6 h

of meetings a week to break even at $14

Part I. The bill

Almost all of it is one vendor

The monthly bill at 100K, 500K or 1M users. It assumes the average user records about 7 hours of meetings a month, heavy, light and free users mixed, with two transcription streams per meeting.

Users

  • Deepgramspeech to text

    $2.02M
    list

  • Claude, GPT, GeminiAI notes

    $123K
    list

  • AWSservers, database, logs

    $100K
    estimate

  • Knocknotifications

    $25K
    list

  • AssemblyAIbackup speech to text

    $21K
    list

  • Amplitudeanalytics

    $10K
    estimate

  • WorkOSauth

    $6K
    list + SSO estimate

  • Sentrycrash reports

    $5K
    estimate

  • Turbopuffersearch (seen in code)

    $4K
    estimate

  • Vercelwebsite

    $500
    estimate

  • Mintlifyhelp docs

    $450
    list

  • Electron, React, Tailwind, Tiptap, Yjs, SQLite, Parakeet, Whisperapp, editor, on-device speech

    Free
    free
  • Total a month$2.31M

At 500K users. Deepgram alone is 87% of it.

About $4.60 per user a month, and 86 to 87% of it is Deepgram.

Big customers usually get discounts, so treat these totals as ceilings.

Part II. Run it yourself

Move the inputs, watch Deepgram

The same model as the table above. Change how many people use it, how many hours they record, what Granola pays Deepgram and which model writes the notes.

Deepgram's price
The model writing the notes

$2.31M

stack bill a month · $4.62 a user

87%

of it is Deepgram

−9%

stack margin on $2.13M revenue

  • Deepgramspeech to text

    $2.02M

  • Claude, GPT, GeminiAI notes

    $123K

  • AWSservers, database, logs

    $100K

  • Knocknotifications

    $25K

  • AssemblyAIbackup speech to text

    $21K

  • Amplitudeanalytics

    $10K

  • WorkOSauth

    $6K

  • Sentrycrash reports

    $5K

  • Turbopuffersearch

    $4K

  • Vercelwebsite

    $500

  • Mintlifyhelp docs

    $450

Our assumptions, not Granola’s: two streams per meeting, about 2% of audio failing over to AssemblyAI, 1 in 4 users paying $17 a seat on average, and per-user estimates for AWS, Knock, Amplitude, WorkOS and Sentry. The app, the editor, SQLite and on-device speech to text are free. Margin is the stack only; salaries come on top.

Part III. The price list

Every service the app calls

Public list prices, with where each one came from. At Granola's size most paid tools move to custom contracts, so every paid price here is a ceiling.

  • The app

    ToolElectron

    List priceFree, open source

    Source (grade)A

  • Screens

    ToolReact + Tailwind

    List priceFree, open source

    Source (grade)A

  • Notes editor and live sync

    ToolTiptap + Yjs

    List priceFree (the open source parts seen in the code)

    Source (grade)A

  • Notes on your laptop

    ToolSQLite

    List priceFree

    Source (grade)A

  • On-device speech to text

    ToolParakeet + Whisper

    List priceFree, runs on your laptop

    Source (grade)A

  • Speech to text

    ToolDeepgram Nova-3

    List price$0.0048 a minute: 29¢ per stream-hour, so 58¢ per meeting-hour (a promo price; regular $0.0077 a minute)

    Source (grade)deepgram.com/pricing (A)

  • Backup speech to text

    ToolAssemblyAI

    List price15¢ per stream-hour, only when Deepgram fails

    Source (grade)assemblyai.com/pricing (A)

  • AI notes

    ToolClaude, GPT, Gemini

    List price0.2¢ (Haiku 5.5) to 11¢ (Opus 5.5) per meeting-hour summary

    Source (grade)Anthropic price table (A) + our token estimate

  • Auth

    ToolWorkOS

    List priceFree to 1M users; SSO $125 (1 to 15) down to $65 (51 to 100) per company

    Source (grade)workos.com/pricing (A)

  • Errors

    ToolSentry

    List priceFree to 5K errors; Team $26 a month for 50K

    Source (grade)saas-calc, 16 Aug 2026 (A)

  • Analytics

    ToolAmplitude

    List priceFree to 2M events a month; Plus to 70M; then custom

    Source (grade)amplitude.com/pricing (A)

  • Notifications

    ToolKnock

    List priceFree for 10K; Starter $250 a month for 50K + 0.5¢ each; Enterprise custom

    Source (grade)knock.app/pricing (A)

  • Search (seen in code)

    ToolTurbopuffer

    List priceFrom $16 a month; Scale $256; Enterprise from $4,096

    Source (grade)turbopuffer.com/pricing (A)

  • Help docs

    ToolMintlify

    List pricePro $450 a month (annual)

    Source (grade)mintlify.com/pricing (A)

  • Website

    ToolVercel

    List pricePro $20 a month a seat + usage

    Source (grade)saas-calc (A)

  • Servers, logs, downloads

    ToolAWS

    List priceUsage based; can't be priced from the code

    Source (grade)n/a

Part IV. One meeting

An hour costs 51¢ to $1.03

Low, mid and high for one hour-long meeting. Speech to text is two streams for the whole hour; the summary is one pass, with chat questions afterwards not counted.

  • Speech to text (2 streams × 60 min)

    Low$0.504 (Deepgram Growth)

    Mid$0.576 (Deepgram pay as you go)

    High$0.924 (Deepgram regular list)

  • AI summary

    Low$0.002 (Haiku 5.5)

    Mid$0.035 (Sonnet 5.5)

    High$0.111 (Opus 5.5)

  • Auth (WorkOS)

    Low$0

    Mid$0

    High$0

  • Errors, analytics, notifications

    Lowunder $0.01

    Midunder $0.01

    Highunder $0.01

  • Total

    Low$0.51

    Mid$0.61

    High$1.03

Against the $14 a month Business plan

  • 2 h

    Hours a month

    9

    Cost at list prices$4 to $9

    Against $14profitable

  • 5 h

    Hours a month

    22

    Cost at list prices$11 to $22

    Against $14around break-even

  • 10 h

    Hours a month

    43

    Cost at list prices$22 to $45

    Against $14loses money at list

  • 15 h

    Hours a month

    65

    Cost at list prices$33 to $67

    Against $14loses money at list

Break-even is 3 to 6 hours of meetings a week. Anyone busier costs more than they pay, so Granola can’t be paying list prices.

13.5 to 27.7 meeting-hours a month. Free-plan users cost the same per meeting and pay nothing.

What the code shows (grade A)

  1. 01

    Two transcription streams per meeting

    One for your microphone, one for the computer's audio, each its own connection (channels: 1). So one meeting-hour is two hours of speech to text.

  2. 02

    Deepgram Nova-3 by default, AssemblyAI as the backup

    The transcription provider defaults to deepgram, with the model nova-3. AssemblyAI takes over if Deepgram fails.

  3. 03

    Silence isn't skipped

    The transcription path forwards everything on both streams. The "skipped audio" counter belongs to on-device speaker separation, not to billing.

  4. 04

    The AI goes through Granola's own server

    Requests go to their llm-proxy and the model is picked there ("Auto"), so we can't see which one. We price a range.

  5. 05

    A private Deepgram endpoint

    wss://<id>.api.deepgram.com/v1/listen, behind the flag deepgram_dedicated_endpoint (off by default). A dedicated setup points to a custom contract, not list price.

Why the real bill is lower

  1. ↓

    On-device transcription

    Parakeet or Whisper on your laptop (opt in, English only). Costs Granola nothing per meeting.

  2. ↓

    A Cloudflare route for Deepgram's models

    Behind cloudflare_transcription_relay_enabled, off by default. A second way to buy the same model.

  3. ↓

    Volume deals

    Deepgram's own pricing page offers Enterprise "significant volume discounts". At Granola's size a large discount is likely, though the amount isn't public.

Part V. How many users?

Nobody outside Granola knows

No official count has been published since 2024, so every figure in this part is an estimate (grade C). This is the evidence the bill's user counts rest on.

  • ~5,000 weekly users, 50% still active after 10 weeks, ~6 meetings a week each (Oct 2024)

    Source (grade)secondary write-ups (C)

    What it suggestsThe only real user figure

  • User base growing ~10% a week (May 2025, CEO)

    Source (grade)secondary (C)

    What it suggestsVery fast growth then

  • Growth since: 4 to 6% a week for two years from 5K

    Source (grade)our maths

    What it suggests~300K to ~2M weekly users. Very sensitive to the rate

  • granola.ai ~2.1M visits a month (Jul 2026); Crunchbase ~1.6M (undated)

    Source (grade)Semrush, Crunchbase (B− / C)

    What it suggestsHundreds of thousands of users, not tens of millions. A desktop app doesn't need web visits, so treat it as a floor

  • ~$30M to $60M a year in revenue (Otter's $500K per employee × ~80 staff; $1.5B valuation at 25 to 50×)

    Source (grade)our inference (C)

    What it suggests~150K to 360K paying seats at $14 to $17 each

  • Named customers: Vanta, Gusto, Asana, Cursor, Mistral and others

    Source (grade)press (B)

    What it suggestsReal company-wide rollouts, not just individuals

Best estimate: roughly 0.5M to 1.5M monthly users, of which ~150K to 350K pay.

Three sizes of company (a model, not a fact)

Inputs that are assumptions: Paid seats; $17 average per seat (a mix of $14 Business and $35 Enterprise); 3 free users per paid user; 12 to 22 meeting-hours a month for paid users and 3 to 5 for free users; an assumed ~50% Deepgram discount; a Sonnet-class summary; other infrastructure at 8% of revenue; ~80 staff (grade B) at an assumed $225K a year fully loaded.

  • Small: 120K paid seats

    Revenue a month (ARR)$2.0M ($24M)

    Meeting-hours a month

    2.5M

    Infra at list

    $1.7M

    Infra with deal

    $1.0M

    Gross margin (deal)

    52%

    After staff (deal)

    −$0.4M a month

  • Mid: 220K paid seats

    Revenue a month (ARR)$3.7M ($45M)

    Meeting-hours a month

    6.4M

    Infra at list

    $4.2M

    Infra with deal

    $2.4M

    Gross margin (deal)

    37%

    After staff (deal)

    −$0.1M a month

  • Large: 350K paid seats

    Revenue a month (ARR)$6.0M ($71M)

    Meeting-hours a month

    13.0M

    Infra at list

    $8.4M

    Infra with deal

    $4.7M

    Gross margin (deal)

    22%

    After staff (deal)

    −$0.2M a month

  • The infra bill is mostly speech to text, about 75 to 90% of it. Granola is a transcription-cost business first.
  • At list prices the mid and large cases lose money on every meeting. That’s why the private Deepgram endpoint and the on-device and Cloudflare routes matter.
  • With a good deal it's roughly break-even after salaries. Granola has raised about $192M, so running near break-even is a choice, not a crisis.

Against real companies

  • Otter.ai: $100M+ ARR, 25M+ users, under 200 staff (Mar 2025)

    Source (grade)Otter's own blog (A)

    What it says about the modelAbout $500K revenue per employee. Granola's ~80 staff × $500K ≈ $40M ARR, close to the mid scenario ($45M). Otter runs its own speech models, so its transcription bill is likely lower

  • Fireflies.ai: ~$1B valuation, said to be profitable since 2023, ~100 to 130 staff

    Source (grade)press and aggregators (B / C)

    What it says about the modelMeeting-notes apps can be profitable. Fireflies raised only ~$19M

  • AI products average ~52% gross margin in 2026; pure application-layer AI ~45%

    Source (grade)ICONIQ State of AI, via secondary write-ups (B−)

    What it says about the modelOur 37% (with a deal) is below average but plausible for an app that rents its speech to text

  • The fastest-growing AI apps run ~25% gross margin, often negative

    Source (grade)Bessemer benchmarks, via secondary write-ups (B−)

    What it says about the modelOur list-price case (a negative margin) is what happens with no discount

So the model is plausible, not proven. The likeliest way it’s wrong is that Granola’s Deepgram rate is far below 50% off (the private endpoint suggests a custom deal), which would push the margin to the ~45 to 55% benchmark range.

Part VI. The limits

What this does not prove

  • ×Granola's actual costs, discounts or margin. None of these are public.
  • ×How many meeting-hours a typical user records.
  • ×Which AI model "Auto" uses for summaries.
  • ×Server, storage and staff costs.

Keep going

That’s Granola’s bill.
Now price yours.

The calculator has 110 tools with verified prices. Pick a budget, pick one tool per row, and see what your stack costs to run each month.

Price your own stack