What does software development with AI agents really cost? A field report with real numbers
In short: I built a fitness and coaching app with a team of AI agents, from the first plan on 17 September to the test version on iPhone and Android on 8 October, and analysed every single request to the models. At list prices, the compute for the whole project would have cost around 16,100 US dollars. The expensive part was not writing code but constantly re-reading the project.
Why I am writing this down
AI agents come with big promises. I found hardly any real numbers from a real project in German, mostly estimates from vendors selling tools. I build websites, shops and automations for my clients and spent the last few weeks building a complete app with AI agents. This is what it really cost and what I would do differently.
The project
- A fitness and coaching app for iPhone and Android: training plans, nutrition, check-ins with feedback from the coach, progress, subscriptions.
- A server with a database and access control, so every client only sees their own data.
- Built by a team of AI agents with fixed roles: coordination, integration, development for interface, data and server, design, quality assurance, security, architecture. I set direction, priorities and approvals, review results and decide everything that concerns law, money and publishing.
- A second AI model from another provider reviews the work independently.
I analysed the entire project from the first commit on 17 September to 8 October 2026, i.e. 22 calendar days, 18 of them with agent work. On top of that comes a detailed analysis of the final sprint week.
The numbers: the whole project
| 17 Sep to 8 Oct 2026 | |
|---|---|
| Duration | 22 calendar days, 18 days with agent work |
| First installable Android version | after 3 days (20 September) |
| Code changes (commits) | 11,825 |
| Merged work packages | 3,041 |
| Reverted packages | 10 |
| Database changes (migrations) | 232 |
| Automated tests | over 11,500 |
| AI agent sessions | 1,116 |
| AI model responses | around 193,000 |
| Tokens processed | 48.2 billion |
| Equivalent at API list prices, main models | around 15,240 US dollars |
| Equivalent at API list prices, second reviewer | around 860 US dollars |
Split by model: Claude Sonnet 5 around 7,090 dollars, Claude Opus 5.5 around 4,270 dollars, Claude Opus 5 around 2,040 dollars, Claude Fable 5.1 around 990 dollars, Claude Sonnet 5.5 around 850 dollars. Models were switched during the project, hence the mix.
The final sprint week in detail
| 5 to 8 Oct 2026 | |
|---|---|
| Test versions (Android and iPhone) | 6 in 4 days |
| Code changes (commits) | around 2,500 |
| Merged work packages | 507 |
| Reverted packages | 2 |
| AI model responses | just under 45,000 |
| Tokens processed | 10.5 billion |
| Equivalent at API list prices, main models | around 3,160 US dollars |
| Equivalent at API list prices, second reviewer | around 360 US dollars |
I do not pay list prices but use subscriptions. The dollar amounts show what the same work would have cost through the providers’ APIs, at their published prices in October 2026. For comparison: development agencies quote ranges of several tens of thousands of dollars and up for a fitness or coaching app. That is not a fair like-for-like comparison, because my own working time, accounts, app stores and legal advice come on top. Still, the order of magnitude is a different one.
Where the costs really come from
This surprised me most: 98.8 percent of all tokens in the whole project were re-reading (98.7 percent in the sprint week). Every agent response sends the conversation so far, the files read and the rules along again. In the sprint week that averaged around 235,000 tokens per response. The text actually written made up only 0.3 percent.
That leads to a simple formula: cost = context size times number of calls. Shorter answers save almost nothing. What does save:
- fewer calls: bundle messages between agents, no acknowledgements like "understood", no waiting loops that check in every few minutes;
- smaller context: restart agents in good time with a short handover instead of letting them run forever; hand searching and reading to helper agents that only report back a result.
In the sprint week the helper agents used only 1.2 percent combined. The most expensive single session, by contrast, kept running with almost twice the intended context and on its own cost as much as several normal sessions.
Which roles use the most (sprint week)
Not the developers, but coordination:
| Role | Share of usage |
|---|---|
| Integration (merges all work packages) | 17.5% |
| Coordination (priorities, alignment with me) | 14.9% |
| Interface development | 11.1% |
| Server development | 11.0% |
| Design | 8.4% |
| Quality assurance | 8.3% |
| other roles | around 29% combined |
Every message to the coordinating roles wakes up a large context. Anyone setting up an agent team should save exactly there.
Full staffing is the most expensive mode
On one day up to eleven agents ran at the same time to finish a test version. That single day cost as much as the two days before it combined. Full staffing pays off for a clear goal with a fixed end. Afterwards I go back to three to five agents.
Second review by a different model
The most valuable decision was a second model from another provider as reviewer. Over the whole project it cost around one eighteenth of the main models and regularly found bugs that the reviewers in my own team had reported as fine: a case where data stayed on a second device after consent was withdrawn, a deadlock in the database, a quantity in millilitres that was calculated as grams. Bugs like these often slip through tests and cost trust later.
The downside: at full load the reviewer’s quota was used up after around 28 hours. The final overall review before the last test version therefore did not happen. Today I keep a fixed share in reserve for it.
What the AI does not decide
- Anything with legal effect: data protection, consents, subscription terms, app store texts. I check these myself and with a lawyer.
- Anything that goes outside: publishing, server changes in production, payments. The agents prepare, I execute.
- Priorities: I decide what gets built next. The agents make suggestions.
By the way, the most frequent type of error was mundane: wrongly guessed times in logs. Agents like to write them from memory. Fixed rules and automated checks catch that.
What I take away
- Agent teams are fast, but not cheap. Running many agents at once without a plan burns in hours what would otherwise last a week.
- Measure costs instead of guessing. I analyse the logs with a small script, in a few seconds and at no extra cost.
- Clear roles and one place that merges. Without that, you get conflicting changes to the same code.
- A second, independent reviewer. It finds what your own team overlooks.
- The human stays responsible. AI agents are a tool, not a substitute for decisions.
What this means for your project
Not every project needs an agent team. For a website, a shop or a single automation, one agent with a clear task is often enough, and costs stay manageable. For larger projects such as your own app or an internal tool, agent teams can save time and money if they are managed properly.
If you want to know whether it pays off for your project: Write to me. In a free initial call we look at what makes sense and what does not.
Free initial call, no obligation. I will tell you honestly whether AI agents are worth it for your project.
Discuss your project