At the end of this part you know what AI-assisted app building costs in credits, which prompts are cheap and which are expensive, and how Git lets you walk away from a failed experiment in two minutes.
Time to read: 10 minutes.
Credits: this part saves you some.
Nobody told me this before I started, so we measured it. Or rather, Claude did: before every step it reminded me to note the start time, and it kept count very precisely. I added the credit numbers from the Copilot usage panel.
The numbers from my two builds
GitHub Copilot bills chat and Agent mode in credits (1 credit = $0.01). I worked on the Free plan, whose allowance is monthly, not per session. Two complete demo apps, the lakehouse portal (about 3 hours) and the forecast app (about 2 h 45 min to scenario planning), used about 50 % of it. By the end of the IBCS polish I was at 81 %.
Then I started measuring single steps:
The lesson: small, precise prompts are cheap; open-ended rounds are not. And layout problems in combined charts are the most expensive kind of fix.
You do: measure your own credits
9a. In VS Code, look at the status bar at the very bottom. Click the small Copilot icon on the right.
9b. The menu shows your usage in percent and the date it resets. Write the number down before a step.
9c. After the step, click the icon again and note the new number. The difference is what that step cost.
For the detailed view and budgets, go to github.com → your profile picture → Settings → Copilot.
What actually drives the cost
Not the time, and not the number of prompts. What the model has to process:
- Every step an agent shows (read a file, edit, run a test, evaluate) is a separate model call. One prompt with 30 steps costs more than several small ones.
- Context is processed again on every call: files read, documentation fetched, the chat history. Long chats get more expensive per step, so start a new chat for each step.
- Fix loops (test fails, fix, test again) multiply the calls.
- Stronger models cost more per call. In the chat input, next to the model picker, „Optimize for“ offers Efficiency, Balance and Intelligence. In Auto mode, Copilot picks the model itself.
Ten habits that saved me credits
- Write precise prompts with exact coordinates (table, measure, ID); every clarification round costs a run. Tip: let your AI companion write the prompt. You bring the facts, it brings the precision.
- Bundle small fixes into one prompt.
- Put the full visual requirements into the first chart prompt.
- Start a new chat for every step.
- Use „Optimize for: Efficiency“ for small fixes; keep the stronger setting for building.
- Add „Do not run browser or UI validation; I will check the UI myself.“ The Data App template otherwise tries to validate charts in a browser, which costs many steps and stops at the Fabric sign-in anyway.
- End prompts with „When you are done, end without further review loops.“
- Stop Copilot when its summary is written; trailing „Working“ phases still run.
- Fix small chart things by hand: it’s Vega-Lite, like Deneb.
- Check usage via the Copilot icon in the VS Code status bar before and after a step.
What the plans cost
As of October 2026 from GitHub’s plans page; prices change often, so check before you start. For customer projects, clarify early who pays for AI credits and whether the customer allows the tool at all.
When to stop: the revert story
In IBCS, variances are often shown as small bars directly above the columns. I wanted that. The first attempt took 6 minutes and 2 points, and the values were right. But stacking two chart views hid the month axis and made the chart wider than its card. Copilot’s fix passed its own tests, but not inside Fabric, and cost another 8 points.
At that point I asked myself: could someone just fix this by hand? Maybe. But not that evening. So I reverted, in two minutes and without a single credit:
You do:
- Source Control: hover over the Changes line and click ↶ Discard All Changes. This drops all uncommitted edits.
- If the failed feature is already in the last commit: git revert HEAD –no-edit. This adds a new commit that undoes the last one. Nothing is deleted, so the experiment can be revisited.
- npx rayfin up and refresh the app.
The idea went to the backlog, with notes for the next attempt. Not every idea is worth implementing at any price. That’s a sentence I say to customers about reports all the time; it applies to AI-built apps just as much.
And this is why I commit before and after every prompt. Without those save points, I’d have spent the rest of my allowance fixing a fix.
Fabric Apps is in preview; menus and commands may change. Written October 2026. Co-authored with Claude as co-learner. Previous: Part 8 · Next: Part 10. #HumanAICoProgramming