The Unit Economics of a $2.99/Month AI Receipt Scanner

Taxr is a receipt scanner. Photo or PDF in, structured expense out: vendor, date, total, tax, category. The incumbents in this category charge $10-31/month. Taxr charges $2.99/month or $29.99/year, and every subscriber is cheap to serve, which is different from the business being profitable. More on that below. This post is the arithmetic, because when we tell people the price, the first response is usually “how?”, and the second is “subscription for a camera app, seriously?”. The answer to both is the same small spreadsheet.

What a scan actually costs

Extraction runs on gpt-4.1-mini. A receipt photo plus prompt averages around 3,000 tokens in. Across a full month of real traffic, blended and with retries included, that works out to about $0.0012 per scan, a tenth of a cent.

There is exactly one model call per receipt. No pre-pass, no second call to categorize; the same response returns the extracted fields and the category.

That makes the per-user maths boring. The median person scanning in a given month does 5 receipts, under a cent of inference. The 95th percentile does about 38 (4.6¢). Our heaviest user last month did 166 (20¢, still 7% of their $2.99). Which is also why the paid plan has no monthly scan cap: at these prices a cap would be a pricing tier dressed up as cost control. Model prices in this class have only fallen since we started building. This business gets cheaper to run every year.

The rest of the stack

Django and Postgres on three shared-CPU Fly.io machines in Sydney. Last invoice: $19.28. Receipt images live on S3, $3.36/month. RevenueCat handles subscription state, free at our size.

Store commission is the biggest line item by far. Apple and Google both take 15% rather than the 30% people assume. The reduced tier exists precisely for small developers, and it applies under $1M a year. Worth knowing that Google applies it automatically while Apple’s Small Business Program is opt-in: it’s a form you have to find and fill in, and it’s easy to leave money on the table without realising. If you ship an iOS app and have never checked, go and look right now.

Net of the store cut, an annual subscriber is worth about $25.49/year against pennies of marginal cost.

Two questions this usually raises, answered directly. Yes, receipt images go to OpenAI’s API for extraction. API data isn’t used for training under their terms, images are stored encrypted on S3, and deleting your account deletes your data (deletion policy). And yes, at the center of the product there’s one model call. “GPT wrapper” is fine by us; the product is the 195-country tax layer, the categorization, the exports, and everything else that makes the result trustworthy enough to hand to an accountant.

Add it up

The whole product runs on about $23/month, all-in: inference, hosting, storage. Eleven annual subscriptions cover it.

The free tier is 10 scans, which costs us about 1.2 cents per free user, ever. That’s the cheapest customer-acquisition budget we could design.

To be precise about “profitable”

The unit economics work: revenue covers infrastructure with room to spare, and each additional user costs pennies. Development time and cost are the real investment, as with every early-stage product. What the spreadsheet shows is narrower and, we think, more interesting: that $2.99/month is a sustainable price rather than a loss-leader waiting to be repriced.

Why the incumbents charge more

They aren’t gouging. They’re selling to different customers.

Dext’s $302/year buys a document pipeline into Xero and QuickBooks with practice dashboards built for accounting firms. Expensify sells approval workflows to teams; its individual plan, around US$55/year, is a side door into expense reports rather than tax records. Keeper and Everlance are US-only and lean on bank-account linking.

The individual freelancer is nobody’s target customer, so the products aimed at her are either firm-grade pipelines or team tooling with the price to match. Those prices are segmentation, not costs. That gap is the entire opening for Taxr.

The honest part

We’re a small independent team and this is an early product. A few thousand receipts scanned all-time, not a few million. At this scale the numbers are easy; the interesting question is what breaks at 1000x.

Our read: inference scales linearly, which is fine. Postgres and S3 are boring, which is also fine. Support is what eventually costs real money. And a $29.99 ARPU means paid acquisition barely works at all, so growth is organic-or-nothing, which is partly why this post exists. Compute was never the scarce resource. Distribution is.

One thing we’d want to know before trusting a small company with tax records, so here it is unprompted: everything exports to Excel and PDF at any time. If Taxr disappeared tomorrow, your records wouldn’t.

The stack, for the curious

React Native and Expo for the app (iOS and Android), Django and DRF on the backend, Postgres, RevenueCat webhooks for subscription state, and gpt-4.1-mini for extraction across 195 country tax configurations: tax labels, fiscal years, currency formats, per-country expense categories. An Australian sole trader sees GST and a financial year that starts in July; a UK freelancer sees VAT and a tax year that starts on 6 April. Honestly more work than the AI parts.

Taxr is on the App Store and Google Play. If you want to see the output before installing anything, the free tools run in the browser.

We’d value your read on two things: where does this pricing break at scale, and what would you want to see before trusting an app with your receipts?