How Pieter Levels made money from an AI game — and what lies behind $1M ARR

Wanted to fly quickly — created his own flight simulator

Pieter Levels' idea was very human: open a game and immediately find yourself in the sky. In his own story he explains that long loading times and updates for Microsoft Flight Simulator pushed him toward a browser experiment. He began development on February 22, 2025, and shared the process on X.

The result is — fly.pieter.com, a multiplayer airplane game. Among the tools listed on its page are Cursor and Three.js: an AI editor helps with code, and a JavaScript library handles 3D in the browser. The author calls the game fully made with AI; we did not conduct a code history audit to independently measure this share.

This illustrates well the idea of Vibe Coding: describing desired behavior, running the result, noticing a problem, and refining the next change. But a generation button does not yet mean that controls, monetization, and support are ready.

Players flew, and businesses bought space in their sky

Monetization had two visible paths. On the game page, verified as of October 19, 2026, the Cessna 172 is available for free, the F-16 is offered for $29.99, and companies are offered in-game advertising. These are active offers, not confirmation of the number of purchases.

The most interesting part is advertising as an environmental element. In the author's WIP journal there are entries about an advertising hot air balloon, the planet Jupiter, banners, and an aircraft carrier. On March 4, 2025, Levels reported selling a Bolt sponsorship for $5,000. This is a specific statement from the author regarding the deal; we do not have bank documents at our disposal.

After the first money, the work did not end. In the entries from March 4–5 — mirror fixes, new missions, radar, self-service ad payment. There is even a reduction in the game night because people didn't like its duration. Behind the loud numbers, you can see typical product work: listening to users and removing inconveniences.

$1M in 17 days? It's important to read these three letters here

MRR usually means monthly recurring revenue, ARR — its annual equivalent. In this case, we do not see the full structure of contracts and cannot confirm that all sales and advertising payments were indeed recurring. Therefore, we retain the author's notation but do not convert it into audited financial statements.

To understand the scale: a $1M annual run rate corresponds to approximately $83,333 per month, if divided mechanically by 12. This is the arithmetic of the metric, not new evidence of actual receipts. Profit requires a different calculation: revenue minus expenses, commissions, returns, and other applicable items. We do not have a verified net profit amount for this game.

A new product from scratch — but the author already had experience

Before this game, Levels created Nomad List, Remote OK, and Photo AI. His own list of projects shows years of launches, including unsuccessful ones. To call him a newcomer who opened a laptop for the first time and got rich would be untrue. The game was new; entrepreneurial experience, public presence, and sales skills already existed.

One notable case does not show what share of all AI games is earning. We see a selected example, not a full sample of launches. The reason for the result cannot be isolated by a single observation either: tools, the idea, public demonstrations, the author's experience, and interest in the new format all acted simultaneously.

What science can actually say about such a story

Scientific research on productivity does not confirm someone else's sales. Instead, it helps test a simple assumption: does AI always accelerate development?

In in a randomized METR study from July 10, 2025 16 experienced developers performed 246 real tasks in familiar open-source repositories. Tasks were randomly distributed between conditions with and without AI access. Using early 2025 tools, work took on average 19% longer. This is a result of a specific sample and type of tasks, not a test of creating a new game or entrepreneurial earnings.

There is also an update from February 24, 2026: METR reported that the new experiment gives an unreliable signal due to the selection of participants and tasks, and the difficulty of measuring work with parallel agents. The authors suggest more acceleration with newer tools but cannot reliably determine its magnitude. It is incorrect to apply the old 19% to all modern Vibe Coding.

Editorial practical conclusion: separately measure time to a working prototype, time for fixes, and the result for the user. And verify willingness to pay separately from the speed of writing code.

What can be repeated without someone else's audience and million-dollar promises

Copying planes is not the same as copying a business. It is more useful to repeat a small validation cycle. For a 3D artist, this could be an interactive preview of their own model or a short browser scene with one clear action. These are our examples of possible experiments, not already proven sources of income.

1. One task. Write down who the user is and what they should do in the first minute. Use the first AI project planner.
2. Small prototype. One scene or function; use existing models and textures. Check performance, loading, and errors.
3. Real reaction. Let a few target people try the result. Note where they stopped, what they went back to do, and what they are already willing to pay for.
4. Clear proposition. If there is demand, define the final result, support, and costs. Our guide to the first order will help separate the demo from paid work.
5. Post-launch check. Count actual payments, returns, costs, and re-use. One viral day does not replace several months of observation.

There is an inspiring part to this story: a small browser game managed to turn attention into a paid offering. There is also a more useful part than the loud headline: the author continued to refine the product and later acknowledged the limits of its sustainability. Both parts are worth seeing before your own launch.