A study, 2026-02-01 to 2026-05-16.

I scanned 30 AI apps. 100% had a hidden cost problem.

30 public AI apps, all shared by their makers on Reddit, Hacker News, Product Hunt, Indie Hackers, and Dev.to. Every one ran through PrePrice's static cost analyzer. Not 80%, not 95% — every single app had at least one high-severity cost issue. The apps are anonymized below; the numbers are exact.

01The Headline
100%

had ≥1 high-severity cost issue

74

high-severity issues in total

2.7

high-severity issues per app (avg)

96%

had a major cost driver we could quantify

This isn't "your AI app might cost too much someday." The trap doors are already in the code, and most makers can't see them. The worst app in the sample loses $269 per heavy user per month at its own recommended price.

Where does your app sit? Find out in a couple of minutes.

Scan a public repo — free, no signup
02The Five Patterns

Five patterns show up in almost every app.

Ranked by how often the pattern appeared across the 30 apps. Each one has a clean fix. Dollar figures are the recoverable spend at 1,000 users per month.

  1. 93%of apps

    Missing prompt caching

    Apps send the same 2,000-8,000-token system prompt on every call. Anthropic and OpenAI both cache repeated input for ~80% off. Most apps never turn it on.

    In the wild. A block-based visual coding tool (hackathon winner) sends four large agent system prompts uncached on every invocation — $168/mo at 100 users, $1,680/mo at 1,000.

  2. 89%of apps

    Agent loops as a top-3 cost driver

    Multi-step agents left uncapped, or capped far too high. If the agent stalls, retries, or loops, the cost runs away with it.

    In the wild. A collaborative whiteboard agent runs its main loop at maxIterations: 50. Capping it lower is worth $420/mo at 100 users, $4,200/mo at 1,000.

  3. 81%of apps

    No rate limits or per-user caps

    The single biggest 'looks fine until it doesn't' pattern. If one user can hit the API 1,000 times in a minute, you don't have a cost model — you have a charity.

    In the wild. A browser-based image and video generator has no rate limit and its API key sits in client JS. A single per-user rate limit is worth $2,790/mo at 1,000 users.

  4. 63%of apps

    Unbounded agent loops, specifically

    The harder version of the loop problem: no cap at all, recursive sub-agents that spawn sub-agents, or 'no limit' written into the docs as design intent.

    In the wild. A permit-review agent (a hackathon winner) forks 5+ parallel frontier-model sub-agents per corrections flow, with no rate limit.

  5. 56%of apps

    Retry amplification

    Fallback chains that retry on every primary failure with no circuit breaker. One flaky upstream call becomes a full second LLM call — sometimes five.

    In the wild. An iOS calorie tracker has fallback retries on by default; every primary failure triggers a full retry on a second provider.

These five patterns are the ones we catch most. See which are in your repo.

Scan a public repo — free, no signup
03The Spread

The median user looks fine. The heavy user explodes the math.

P95 / P50 is how much worse the heavy user is than the median. The eight apps with the widest gap. If your dashboard shows averagecost per user, you can't see this until churn or your Stripe balance tells you.

App (anonymized)P50 / user / moP95 / user / moMultiple
An iOS calorie tracker$3.42$16.824.9×
A voice-to-text dictation tool$0.12$0.544.5×
A multi-provider chatbot platform$9.24$36.964.0×
A diabetes companion app$2.83$11.324.0×
An open-source meeting notetaker$0.02$0.084.0×
A sovereign autonomous agent$8.42$33.684.0×
A desktop cartoon studio$22.85$91.404.0×
An AI coding IDE$0.87$3.484.0×

PrePrice prices your P50 and P95 separately, so the heavy user is visible before launch.

Scan a public repo — free, no signup
04Underwater

Five apps lose money on a heavy user at their own price.

Of the 24 apps with a recommended price, five are underwater at P95. One is underwater at P50 too — its median user already costs 9× the sticker price.

App (anonymized)P95 cost / userRecommended priceLoss / heavy user
A browser image + video generator$279.00$9.99−$269/mo
A desktop cartoon studio$91.40$49.00−$42/mo
An iOS calorie tracker$16.82$9.99−$7/mo
A collaborative whiteboard agent$37.02$29.99−$7/mo
An AI shorts generator$44.58$39.00−$6/mo

Before you pick a price, know what a heavy user actually costs you.

Scan a public repo — free, no signup
05The Pricing Problem

Pure subscription pricing rarely works for AI apps.

For every app, PrePrice computed a sustainable pricing model. A handful of heavy users eats all your margin on a flat subscription — you need a usage cap or a per-call charge somewhere. The distribution across the 30 apps:

  1. 13

    Hybrid

    subscription + usage + free tier — the modal recommendation

  2. 6

    Freemium

    subscription with a generous free tier

  3. 4

    Per-invocation

    pay-per-use

  4. 4

    Pure subscription

    flat monthly

  5. 3

    Usage-based

    metered

Every scan ends with a recommended pricing model tuned to your cost shape.

Scan a public repo — free, no signup
06The Totals

$31,000 a month of fixable waste, sitting in the code.

Combined recoverable spend across the 27 scanned apps, at 1,000 users each. Per app, that's roughly $1,150/month left on the table — from fixes that each take about 30 minutes.

At 100 users each

$3,102

per month · $37,228 / year combined

At 1,000 users each

$31,024

per month · $372,282 / year combined

At 10,000 users each

$310,235

per month · $3.7M / year combined

The scan is free. Finding one $400/month leak pays for the audit ten times over.

Scan a public repo — free, no signup
07Methodology

The sample. 30 AI apps shared publicly by their makers on Reddit, Hacker News, Product Hunt, Indie Hackers, and Dev.to between 2026-02-01 and 2026-05-16. Filter: a public GitHub repo, uses a paid LLM or AI API (OpenAI, Anthropic, Gemini, Replicate, fal.ai, ElevenLabs, etc.), and shared by the maker themselves. 27 of the 30 produced a clean scan; the aggregate statistics are computed over those 27.

Static analysis only.No runtime profiling. Costs are derived from model defaults, system prompt sizes, agent loop caps, retry policy, and stated usage assumptions (10 actions per user per day). For agent flows where per-turn tokens couldn't be detected, PrePrice used 2,500 input / 1,200 output tokens as a default.

Pricing source.Anthropic, OpenAI, Google, and provider pricing pages as of May 2026. Recommended prices are PrePrice's computed suggestion, not what each app actually charges — most of these apps are pre-launch or open-source and have no stated price yet.

Sample bias. These are apps shared publicly with code. They skew toward indie developers, hackathon winners, and side projects. Closed-source paid SaaS is not represented.

Anonymization.The aggregate statistics on this page were published in May 2026. Per-app figures were not, so every app here is described by category, not named. If you recognize your own app and want the full, named scan output, it's a scan away.

Run the exact same analysis on your repo. No signup, nothing stored.

Scan a public repo — free, no signup

Find out where your app sits in this study.

Paste a public GitHub URL and get the same analysis that produced these numbers — per-user cost at P50 and P95, your top cost drivers, and a price recommendation. No account, no email, no card. The clone is wiped as soon as the report is written.