Most answers to this question are guesses. Ours is measured: we host AI-built apps and meter every model call, so here is the actual distribution across 570 apps that were active on our platform in the last 60 days (aggregate platform telemetry, August 2026, rounded):
| Metric | Model cost |
|---|---|
| Median total AI spend per app | $0.17 |
| Average total AI spend per app | $0.50 |
| 90th percentile app | $1.32 |
| Median single request to the agent | $0.036 |
| Typical first build of a small app | $0.04 to $0.09 |
Read that again: the median app has consumed seventeen cents of AI. The 90th-percentile builder, someone who iterated seriously for weeks, spent about the price of a coffee. The averages are pulled up by a small number of heavy builds, not by any per-app floor.
Why so low, and what the catch is
Two reasons the raw model cost is small. First, modern coding models are cheap per request: a real feature request ("add client logins with saved bookings") costs a few cents of tokens. Second, most of an app's code is written in the first handful of requests; iteration edits files rather than regenerating them.
The catch is that model cost is not the whole bill anywhere:
- Hosting is the recurring cost. A deployed app needs a server, a domain, and storage, whether that is a platform subscription, a VPS, or serverless bills. This is usually the largest line item within a few months.
- Platform plans bundle the two. Chat-based builders typically sell a monthly plan that includes model credits and hosting together, in the $20 to $50 range. That is the realistic "cost of building an app with AI" for a non-developer: tens of dollars a month, not thousands up front.
- Third-party services (payments, SMS, maps) bill you directly on their own accounts, same as they would a human developer.
Bring-your-own-key changes the math
If you already pay Anthropic or OpenAI directly, some builders let you plug in your own API key, so model usage bills at raw provider prices with no markup. On our platform, BYOK calls route straight to your provider and we add nothing; the plan then only pays for hosting and infrastructure. We published the real per-call arithmetic and which builders actually support BYOK separately.
Against the alternatives
| Route | Typical cost | What you get |
|---|---|---|
| Agency or dev shop | $10,000 to $150,000 | Custom software, months of lead time |
| Freelancer | $2,000 to $20,000 | Custom software, weeks |
| Classic no-code (drag-and-drop) | $25 to $100/mo + your weeks of layout work | An app shaped like the template |
| AI builder | $0 to ~$50/mo | A working custom web app in hours, iterated by chat |
The honest caveat on the last row: the ceiling is real. An AI builder gets a booking system, a tracker, a storefront, a dashboard, a game, live this week. It does not replace an engineering team for a venture-scale product, and native App Store distribution remains its own project on any tool.
What actually drives your cost up
From our fleet data, the expensive apps are not the ambitious ones; they are the vague ones. Ten rounds of "that's not quite right" costs more than one precise request. Screenshots of what you want, concrete copy, and one change per message keep a serious build inside a dollar of model spend. The p90 number above, $1.32, includes people who iterated daily for weeks.
The practical bottom line
If someone quotes you four figures to "build an app with AI," the AI is not what you are paying for. The model work for a typical small-business app costs cents; the recurring costs are hosting and whatever plan bundles it. You can verify this for free: VibeKit's free tier builds and deploys your first version at $0, and you will have spent under a dime of anyone's compute finding out whether the idea works.