Demolition Cat — From Rental Flat to Museum

A cat, two prompts, one invoice

An orange cat is left alone in a flat with 90 seconds to wreck it: push over the furniture, smash the crockery, collect tins of cat food, spend the tins on upgrades that unlock nicer rooms. The first level is a cheap rental. The last one is a museum.

The game is the demo. The number worth reading is the invoice: two prompts, 17 model calls, $5.77, and not one line of code typed by hand. The model was Fusion Max 2.0 from Singapore's FUMO Lab, reached through an OpenAI-compatible endpoint, so the same loop runs from Cursor or Codex once you fill in a base URL and a key.

The first prompt is a specification, not a wish

Most "make me a game" prompts fail because there is nothing to fail against. This one starts by asking for a design document, and the document is judged against CrazyGames' own evaluation criteria: sessions of at least 10 minutes, day-2 retention between 10% and 15%, playable within three seconds, everything generated in code so there is no asset to download.

That reframing does more work than the choice of model. A platform's review metrics are a free specification, and they force the loop the game needs — short runs, upgrades, daily tasks, a reason to come back — before any code exists. The document came out at roughly 27,000 characters; the build that followed was about 130 KB and ran on the first pass.

Four rounds, sixteen of twenty

Four revision rounds followed. Twenty requirements were raised across them and 16 were confirmed; the author says the other four needed another pass. None of the fixes were typed by hand, which is the part worth noting: each round was a new prompt against the previous build, not a debugging session.

The final pass was art, and it is the one step that leaves the model. The placeholder cat became a Blender model: an orange tabby with 24 layers of fur.

What the cost comparison actually measures

$5.77 for 17 calls. Priced at Claude Fable 5.1 or GPT-6 Astra rates, the same volume comes to about $14.92, so the ratio is 2.6×. Across 100 prototypes the difference is roughly $900.

The dollar figure is real — it is a measured bill, not an estimate. Two things sit underneath it. First, it depends on new-user trial credits and on the vendor's current promotional pricing, so it is the cost of this attempt today rather than a stable unit price. Second, the article carries a referral link to the model vendor, which makes the $14.92 comparison a vendor-side calculation rather than a neutral benchmark.

Still, the argument does not depend on the exact ratio. Solo developer anul agarwal made a Suika-style variation with ChatGPT in 2023 that reached 65M+ players and earned more than $25,000, and he is explicit that it took 120+ prototypes and the first 100 barely earned anything. If one attempt costs $5.77 instead of $15, that path stops being reserved for people with a budget. If it costs $5.77 and the first 100 still earn nothing, the cheaper loop buys you more attempts but not a better hit rate.

Nothing here has shipped

The game is still on the author's machine. It has not been submitted to CrazyGames or to WeChat's mini-game platform, so it has no players, no ad impressions and no revenue.

What is left is the part that models cannot do for you: CrazyGames runs three review lines, WeChat mini-games come with filing requirements, and neither is a design problem. The author's own closing note is that most small games still will not land, however cheap the build was.

What transfers

The reusable asset here is not the game and not the model — it is the prompt order. Write the spec against the platform's metrics first, then bound the iteration to a fixed number of rounds so the cost per attempt stays known.

Everything else is conditional. The cost advantage assumes a cheap endpoint and a small, self-contained build. The art still needs a human and a tool that is not a language model. And the two prompts are published in full, which is the reason this belongs in the case library: you can run the workflow against your own idea instead of taking the conclusion on trust.