Vibe-coded apps have a reputation for looking right and breaking on the second tap. These two do not. We gave four AI models two app briefs, a phone finance app and a team kanban board, one prompt each, and let them build without anyone fixing the result. Tap through the finance app's send flow or drag a card across the board: it works.
What makes these vibe-coded apps work?
The prompt made the data and the flow concrete, and left the design to the model. The finance brief supplied the balance, transactions and contacts as a file and described the four-step send flow; the kanban brief supplied the cards, people and columns and listed the interactions. With that much fixed, every model delivered an app whose numbers add up: after sending $42.50 each finance app shows $4,783.69, down from $4,826.19, with the transfer at the top of the activity list.
What did each model's apps feel like?
Recognisably different, even with the same data.
- Claude Opus 5.5's finance app vibrates the phone on key taps and ends the flow on a full-screen green "$42.50 sent"; its kanban board tracks "3 of 14 done" in the header and tilts the card you drag. 69 and 101 minutes.
- Claude Fable 5.1's finance app animates the balance counting down after sending; its kanban board is the most compact, with overdue cards outlined in red and a delete option in the details panel.
- GPT-6 Astra's apps are airy and calm, with a numbered step indicator and a sidebar card that reads "Small steps. Good things ahead."
- GPT-6.1 Sol's finance app draws a fake iPhone status bar and confirms with a slide-up sheet; its board's sidebar card reads "Good things are growing."
What would it take to ship one?
More than a prompt. These apps run entirely in the browser from a single file with fixed data: no accounts, no server, no real payments. The kanban boards save to your browser's local storage, which is fine for a prototype and not for a team. Treat a vibe-coded app as a very good, clickable first draft, and plan the backend, authentication and security review separately, as the OWASP guidance for web applications recommends for anything that handles real data.
How do you vibe code an app that works?
Fix the data and the flows, free the design. Our two briefs, the finance app and the kanban board, are public and make good templates: supply your data as a file, list each interaction step by step, require a phone layout, and let the model choose the look. Then click through every step yourself before you show anyone.
Four boards, names hidden. Drag a card in each and pick the one that feels right.
Take the blind test
Frequently asked questions
Do vibe-coded apps actually work?
These do. Every finance app here completes a four-step send-money flow and updates the balance correctly, and every kanban board supports drag and drop between columns, undo, filters and a details panel, with no runtime errors in our recordings.
How long does it take to vibe code an app?
Between 9 and 69 minutes for the finance app and 12 to 101 minutes for the kanban board, depending on the model. The GPT models were fastest; the Claude models took longer and added more polish.
Are vibe-coded apps safe to use for real money?
No. These are front-end apps with a fixed dataset and no backend, built to compare models. A real finance app needs authentication, a server, payment rails and security review that no single prompt provides.
What is the best AI for vibe coding an app?
For the most polished first attempt, a Claude model; for fast, cheap attempts you can repeat, GPT-6.1 Sol. Our app-development comparison goes into the details brief by brief.