Case study · On-device AI · Android & iOS

At 100,000 users Offhand would pay at least $31K a year for cloud AI. But it pays zero. How?

Real usage from my voice notes app on Android and iOS, today's API prices, and the math of running everything on the phone.

By Dmytro Samoilov

From $31K a year cloud transcription bill at 100,000 users that Offhand doesn't pay

Every 10-minute voice note that goes through a cloud speech-to-text API costs somebody 2.5 to 6 cents.

In Offhand it costs nothing.

What Offhand is

Offhand is a voice notes app. You record your voice note and a couple of minutes later there is a readable note with a title and sections from it. All with on-device AI.

There are two native apps, one for Android and one for iOS. The logic is shared with Kotlin Multiplatform, and both apps run the same two open-source models right on the phone. Whisper small turns speech into text, and Gemma 4 E2B turns that text into a note. The phone downloads them once, around 3 GB, and after that nothing leaves the phone.

So I'm not buying anything. No API, no server, no GPU in somebody's data center, on Android and on iOS.

I built it this way first of all for myself. I didn't feel good about my voice going to some server, because in those recordings there are names of my clients, my plans, sometimes my family. The money part came as a side effect, and that's what this article is about.

My real usage data

Offhand has hundreds of active users. The app sends anonymous usage stats from people who share them, only events like "a note was recorded" and never the content.

In the last 28 days people who recorded made about 3.5 notes each.

The average note length I don't have yet. The app sends it, but my reports don't show it yet, and I don't want to guess one number. So I calculate three cases:

What the cloud charges

Public pay-as-you-go prices for batch transcription, per minute of audio (as of October 25, 2026):

I take the cheapest and the most expensive one from this list, $0.0025 and $0.006, and do all the math with both.

The summary step I leave out. On Gemini 3.1 Flash-Lite an hour of speech costs about 2 cents to summarize, so transcription is the real bill.

The math

A 10-minute note is 2.5 cents on the cheapest API and 6 cents on the most expensive one.

One user makes 3.5 notes a month. With 10-minute notes that's 35 minutes, so 9 to 21 cents per user per month. Basically nothing for one person.

Per year, at 100,000 users who record:

At 10,000 users divide by ten. At 1,000,000 users multiply by ten, so with 10-minute notes that's $1.05M to $2.52M a year.

In Offhand all of this is zero, on both platforms.

The 100,000th user costs me the same as the first one.

On the other side, a cloud bill could be lower in two ways. At 100,000 users you would negotiate, Deepgram's Growth plan is already $0.0036 and Google's standard rate drops after 500,000 minutes a month. And the cheapest route I found is sending the audio straight to Gemini 3.1 Flash-Lite, about $0.0013 a minute, roughly half of the cheapest API above. I didn't test its transcription quality personally though.

Prices also move both ways. Gemini 3.8 Flash for example doubles its list price on January 1, 2027. So I would read all of this as an order of magnitude and not as an invoice.

What I pay instead

Zero per user doesn't mean zero cost, the cost just moved.

The phone needs 5 GB of RAM and space for 3 GB of models, so not every phone can run it. On Android 14% of the phones that install Offhand don't pass the check and 86% do. That's the price on the local side, a bit smaller market and the engineering time to make it work, which I paid once and not per user.

And it doesn't have to be all or nothing. You can run on the phone where it's strong enough and keep the cloud for the rest. With an install base like mine, and if people use the app more or less the same on every phone, roughly 86% of the minutes would leave the bill.

Offhand is available for Android and iOS. Repo is public, so you don't have to take my word for it.

Prices from AssemblyAI, OpenAI, Google Cloud, Deepgram, AWS and Gemini, checked on October 25, 2026. Usage from Offhand anonymous analytics, 28 days up to October 25, 2026.

Want this math for your app?

If your AI bill grows with your users, or you need on-device AI working in production on Android or iOS — that's what I do. AI cost audits and on-device / hybrid architecture for mobile teams.

Book a call