There are two costs, and people only ever ask about the first one. The build is what you pay to get the thing to exist. The running cost is what you pay every month for as long as anyone uses it, and it is the number that decides whether your idea is a business.
Cost 1: getting it built
| How | Rough cost for a first usable version | What you are really buying |
|---|---|---|
| You, with an AI coding assistant | The subscription, plus your evenings | Full control, a slow start, and you learn what the thing actually needs |
| A freelance developer | Low thousands | Speed, and a dependency on one person's availability |
| An agency | Tens of thousands | A team, a contract, and a process that assumes you know what you want |
The ranges are wide because "a first usable version" means whatever you decide it means. The single biggest lever on this number is not the developer's rate. It is how much you ask for. Half the features in a typical first brief are never used.
If you are considering the first row, how to install Claude Code walks through the setup from zero.
Cost 2: what it costs every month
This is the part that surprises people, so here it is in order of how likely it is to hurt.
AI usage, charged per use
AI providers charge by the token. A token is roughly three quarters of a word. You pay for what you send (the customer's text, plus the instructions, plus any documents you attach) and for what comes back, with the reply costing several times more per token than the input.
Put plainly: a short, simple AI feature costs a fraction of a cent per use. A long one that reads documents and writes a page of output costs somewhere between a few cents and a couple of euros.
Multiply by how often each customer triggers it. That is the number that matters:
| If one use costs | and a customer uses it 50 times a month | then per customer, per month |
|---|---|---|
| €0.002 | 50 | €0.10 |
| €0.05 | 50 | €2.50 |
| €0.60 | 50 | €30.00 |
If you charge €19 a month and sit in the last row, every new customer loses you money, and growth makes it worse. This is the single most common way AI products fail commercially, and you can work it out before writing any code.
Hosting
A small application with a few hundred users runs happily on a virtual server for roughly €5 to €20 a month. Managed platforms that deploy from a repository cost more but save you evenings. This is rarely the line that kills you.
Payments
Payment providers take a percentage plus a fixed fee per transaction, typically a few percent. On a €15 product that fixed fee matters more than the percentage, which is an argument for selling in slightly larger units.
The small ones that add up
- Domain name: €10 to €20 a year.
- Transactional email, so your messages do not land in spam: free at low volume, then a few euros a month.
- Error tracking and uptime monitoring: free tiers are genuinely enough at the start.
- Your own time answering customers: the most expensive item on this list, and the only one with no free tier.
The costs nobody warns you about
- Failed runs. When an AI job fails halfway and retries, you pay for both attempts. Three failures in a row on a paid order is real money for zero delivered value.
- Curious people and bots. Anything free and public will be hit by scripts. Without a limit, a single afternoon can cost more than a month of normal use.
- Web search inside AI features. Letting the model browse multiplies the input it reads, and providers charge separately per search. It can triple the cost of a feature, so switch it on deliberately.
- Refunds and disputes. A disputed payment costs you the amount plus a fee.
A worked example
Say you sell a report generated by AI, one payment per report, €15.
| Item | Per report |
|---|---|
| AI generation | €0.60 |
| Payment fees | €0.70 |
| Email delivery | €0.00 |
| Share of hosting at 100 reports a month | €0.15 |
| Total | €1.45 |
That leaves €13.55 per sale before your time and before marketing, which is a healthy shape. Now change one assumption: give the free version an AI feature that costs €0.20 and have a hundred people try it for every one who buys. The free tier now costs €20 per sale, and your €13.55 margin is gone.
The lesson is not "no free tier". It is that the free part must be the cheap part, and there must be a cap.
Five decisions that keep you solvent
- Work out the cost per use before you build. A spreadsheet with three numbers beats a surprise invoice.
- Set a hard limit on spending in your provider's dashboard. Every serious one supports it.
- Make the expensive part happen after payment, not before.
- Put a human check in front of anything public and free, so scripts cannot run up your bill.
- Charge more than you think. Doubling a price is easier than halving a cost.
Know which parts are expensive before you start
All of the above assumes you know which features actually need AI and how often they run. If you are not sure, that is the thing to settle first: does your app idea actually need AI takes it apart feature by feature, and the free check on this site does the same for your specific description, including running-cost scenarios in the full report.