AI Trends

Making the 3D web mini-game "Cat Demolition" with Fusion Max 2.0 for $5.77

Ayi AI NotesUpdated Nov 12 21:51

AI summary

The author used FUMO Lab's Fusion Max 2.0 to build a playable 3D web mini-game from scratch, "Cat Demolition: From Rented Apartment to Museum," without writing any code by hand, spending a total of $5.77 from design and coding to four rounds of revisions. Calculated at official public pricing for the same usage, switching to Claude Fable 5.1 or GPT-6 Astra would cost about $14.92, and for 100 prototypes the difference would be over $900. The article also provides two prompts, a 4-step integration method, and the publishing requirements for CrazyGames and WeChat mini-games.

Why read this

AI score 58

The author fully discloses the prompts and cost breakdown from the design document through four rounds of iteration, which can be used to estimate the trial-and-error cost of a mini-game prototype.

Topics

Tags

Body · AI translation

Independent developer Anul Agarwal used ChatGPT to write code and built a Suika Game variant, earning over $25,000 in total. But he also said he made more than 120 prototypes along the way, and the first 100 barely made any money. For a small game to make money, what matters is being able to keep trying, and every try comes with a model bill.

This time I didn't write a single line of game code. Using Fusion Max 2.0, the fusion model from Singapore's FUMO Lab, I made a playable 3D web mini-game called "Cat Demolition: From Rental Apartment to Museum."

Let's look at the finished product first. What's surprising is that this video was also made purely with Fusion Max code:

From writing the design and writing the code to four rounds of revisions, it cost only $5.77. Calculated at official public pricing for the same usage, switching to Claude Fable 5.1 or GPT-6 Astra would cost about $14.92, 2.6 times as much; for 100 prototypes, that's a difference of over $900.

This article lays out the whole process: the original text of both prompts, what each of the four revision rounds changed, the cost breakdown of 17 calls, the 4 steps to integrate it, and what hoops you have to jump through to get onto CrazyGames and WeChat Mini Games.

If you want to try it yourself, new FUMO users get trial credits upon registration, which at input pricing works out to roughly 1.67 million tokens of Fusion Max 2.0, or 2.5 million tokens of Fusion Code 1.0. The registration link is here: https://fumolab.ai/en?utm_source=9dta63zy4y .

In this article we discuss:

▸ Trial-and-error cost · Over 120 prototypes, the first 100 barely made any money ▸ Model selection · Why Fusion Max 2.0: 128K output per call, tokens from internal collaboration aren't billed ▸ The finished product first · From a 27,000-character design document to 130KB of first-version code ▸ Four rounds of revisions · 16 of 20 requirements confirmed, not a single line of code hand-edited ▸ Swapping in a Blender fat orange cat with 24 layers of fur ▸ The cost ledger · $5.77 to make it playable, about $14.92 with another provider, over $900 difference across 100 prototypes ▸ Try it yourself · Claim trial credits, 4-step integration, copy both prompts verbatim ▸ Publishing · CrazyGames' three assessment thresholds, the four materials WeChat Mini Games requires, and a bucket of cold water

1. Trial-and-error cost · Over 120 prototypes, the first 100 barely made any money

The Suika Game variant mentioned at the start was made by Anul Agarwal in 2023, writing code with ChatGPT and drawing images with DALL-E; he wrote about it this May. He put it even more bluntly in another post: about 6 years of making games, over 120 prototypes in total, the first 100 barely made any money, and only the roughly 20 after that came closer to six figures.

What I take from this is that for a small game to make money, you have to be able to keep trying at low cost. Every prototype comes with a model bill, and the lower the bill, the more tries you can afford. So the one thing I wanted to test this time was: using Fusion Max 2.0 to make a playable game from scratch, how good is the result, and how much does it cost.

2. Model selection · Why Fusion Max 2.0: 128K output per call, tokens from internal collaboration aren't billed

Fusion uses a self-developed routing model to pick the right model for each step of a task; at key steps, it has multiple models analyze from different angles and then synthesizes the conclusions into one answer. Traditional routing generally picks just one model at the request entry point, which is the biggest difference between the two.

Applied to this game, three points line up perfectly:

It has to spit out an entire game file at once. What I wanted was a single-file index.html, fully rewritten on every edit, and this time each round's output was between 45,000 and 64,000 tokens. Max 2.0 has a 1M-token context and outputs up to 128K tokens per call, enough to hold the whole file plus a set of revision notes.

It has to think it through and write a lot. First write 27,000 characters of gameplay and numerical design, then write code according to the design and fix code according to bugs, which is exactly what Max 2.0 focuses on: web programming and multi-turn Agent tasks.

The more revisions, the more the bill has to stay under control. The official documentation states that tokens generated by internal multi-model collaboration aren't billed; the bill is calculated only on the usage returned by the API.

Whether it's actually capable comes down to what it writes in practice.

3. The finished product first · From a 27,000-character design document to 130KB of first-version code

First, here's some actual gameplay footage of the final game,

This is "Cat Demolition: From Rental Apartment to Museum," which I made with Fusion Max 2.0. The player controls an orange cat and, within 90 seconds, causes as much destruction as possible—knocking over furniture, smashing objects, and earning cat cans, which are then used to upgrade attributes and unlock new scenes. The game has 3D scenes, character control, destruction scoring, an upgrade system, and saves, and it can be played on both computer and phone.

However, the results in the video didn't come from the first generation. In my view, no matter which model you use, expecting to get a completely satisfactory finished product from a single generation isn't very realistic. This result was polished bit by bit through four rounds of revisions and multiple playtests.

Let me start with how it was done at the very beginning. What I asked Fusion Max 2.0 for was: a single-file index.html, Three.js loaded via CDN, no external images or 3D models, the room and the orange cat all generated in code as low-poly, playable on both computer and phone, with a 90-second countdown, destruction scoring, cat-can upgrades, and 4 unlockable scenes.

In the first step, I gave the gameplay requirements to Fusion Max 2.0, set the reasoning intensity to high, and got a 27,000-character design document that included the numerical formulas, prop lists for the four scenes, feel parameters, and save structure.

In the second step, I handed the design document back as-is and had it write the complete code according to the document, which produced a 130KB index.html.

When I opened the first version, the start screen was already fairly complete, and the countdown and daily task bar were there too. After actually entering the game, there were still several areas that needed further polishing:

Scene rendering: about 1 to 2 seconds after clicking start, the 3D scene would disappear, leaving only the background color on screen;

Interaction details: the click on start was also recognized as a paw swipe, so the game began with 18 extra demolition points;

Character design: the first version of the orange cat was a fairly simplified low-poly model, and its outline and proportions weren't ideal yet, blocking part of the view.

The first version had already set up the game's basic framework, and scene stability, interaction, and character presentation were left to later rounds for polishing.

Below are two images: the screen after clicking start in the first version, and the first-version orange cat after the first round of rendering fixes.

4. Four consecutive revision rounds · 16 of 20 requirements confirmed, not a single line of code changed by hand

The revision method was simple: I playtested in headless Chrome, wrote the problems I saw into a paragraph, handed it back together with the entire current code, and asked for a complete new file. After four rounds, I hadn't changed a single line of the game code by hand; the only human work was playtesting, describing problems, and taking screenshots.

Round one: the disappearing scene and the accidental paw swipe were fixed in one round. I only described the symptoms, and it located the causes itself. The cat's design still wasn't right in this round and was left for the next round.

Round two: the first-version design was replaced with a low-poly orange cat with ears, stripes, four legs, and a tail. The main menu, upgrades, scene gallery, and converting an interrupted match into cans all worked, the console showed no errors, and it held a stable 60 frames. The camera still needed fine-tuning: the previous round had added boundaries to the camera to avoid filming outside the room, and the cat would go off-screen when it walked to the edge.

Round three: the camera could now follow the cat steadily. I used a script to play a complete 90-second match, and the settlement rating was Furniture Assassin, with 394 demolition points, 55 cans earned, and 6 new gallery entries unlocked; upgrades, scenes, daily tasks, and saves all worked. What remained to optimize was the camera distance: on phone the cat looked too large, and when it was against a wall, part of the view was blocked by the wall.

This is the settlement card from that match.

Round four: the camera pulled back so most of the room was visible; the wall on the camera side would automatically hide itself; and the on-screen control hint on phone was changed to holding down the paw-swipe button. By this point, of the 20 requirements I had listed in advance, 16 could be directly confirmed by script, while particles, screen shake, slow motion, and sound effects had to be checked by eye and ear.

After switching to the Blender-made cat later, the same rental apartment's desktop and phone versions look like this.

Overall, Fusion Max 2.0 understood bugs described in plain language quite well, and most problems could be fixed correctly in one round. In each of the four rounds it had to rewrite the entire roughly 130KB of code, and I ran regression playtests every round; existing features remained stable throughout, with no features lost.

Of course, making a complete game inherently means testing and revising as you go, and individual adjustments can bring out new detail issues, so it's worth playtesting again after every round. From reporting problems to getting the code fixed, this back-and-forth iteration is something Fusion Max 2.0 can handle.

5. Switching to a Blender chubby orange cat and 24 layers of fur

The low-poly cat felt a bit stiff. Instead of using a third-party modeling platform, I had Fusion Max 2.0 write a Blender 5.2 Python script, run it headlessly on my machine, and export a GLB model with skeletal animation.

The first version of the script ran directly in Blender without errors, but the cat was too thin and its color was grayish. After four more rounds, it gained half-closed eyes, round cheeks, a white muzzle, and a chest fur tuft.

Running fur in the browser was another challenge. It wrote a separate preview page using 24-layer shell fur, and at 60 frames it already felt like a plush toy up close.

The three images below are the chubby orange cat in Blender, the 24-layer fur preview, and a close-up after it was put into the game.

When merging the cat back into the game, one detail left a deep impression on me: the fur effect in the game differed somewhat from the preview page. I asked it to keep optimizing, and instead of directly touching the code, it first proposed that it needed the fur preview page code and comparison screenshots to pinpoint the problem more accurately. When context is missing, it points out the information gap first rather than blindly making changes.

After the information was filled in, the close-up effect improved further—the facial outline, white muzzle, and stripes were all clearly visible. At the camera distance of normal gameplay, the fur's fluffiness still falls a bit short of the preview page, but the character as a whole is much more complete. Web games have to balance real-time rendering and a smooth 60 frames at the same time, and fine fur is inherently a hard job—there's still room for further optimization down the road.

6. The cost ledger · Getting it to a playable state cost $5.77; with another provider it would be about $14.92, and 100 prototypes would differ by over $900

Let's start with unit prices. In the standard tier with input not exceeding 200K tokens, Fusion Max 2.0 is $3 per million input tokens, $0.30 for cached input, and $20 for output. Claude Fable 5.1 and GPT-6 Astra are both $10 for input and $50 for output. For a task like writing code, where output far exceeds input, the output unit price determines how thick the bill is—$20 versus $50, a 2.5x difference on this item alone.

The FUMO column is the usage returned by the API multiplied by the official public prices. I don't have keys for the other two, so I didn't actually run them; it's the same usage multiplied by their official public prices, which is an estimate. If they actually run with longer reasoning or need more revision rounds, it would only be more expensive. Another 3 calls didn't return usage and are not included in the table.

Converted to prototypes: based on the usage it took to reach a playable state this time, 100 prototypes would cost about $577 on FUMO, and about $1,492 estimated at Fable 5.1 or Astra unit prices. For an indie developer, this nearly thousand-dollar difference could be exactly the travel fare for making it to the next hit.

7. Try it yourself · Claim trial credits, integrate in 4 steps, copy the two prompts verbatim

For those interested, I suggest first claiming FUMO's new-user perk: converted at the input price, it's roughly equivalent to 1.67 million tokens of Fusion Max 2.0, or 2.5 million tokens of Fusion Code 1.0. Whether for making a small game or testing your own AI coding project, it's worth a try. Registration link: https://fumolab.ai/en?utm_source=9dta63zy4y

Follow the official getting-started docs to integrate:

Register and log in to the FUMO console;

Open API Keys, click Create New Key, and create a new key;

Set the Base URL to https://capi.fumolab.ai/api/fusion/v1 and the model name to fusion-max-2.0;

Verify the API is reachable with one command:

It's compatible with OpenAI's Chat Completions API; tools like Cursor and Codex can use it once you fill in the Base URL and key. Billing only counts the usage in the response; if a reply gets truncated midway, don't use it to overwrite the previous version of a file that still runs.

The two prompts used this time are as follows, verbatim:

First prompt | Gameplay and numerical design:

Second prompt | Write code according to the design:

The four subsequent revision rounds had no general template; each round was the problem I saw at the time plus the current full code.

8. Publishing · CrazyGames' three assessment lines, the four materials WeChat Mini Games require, and a bucket of cold water

Once the code is written, if you want to make money on overseas platforms, you still have to pass the platform's testing period first.

Take CrazyGames as an example: new games first enter the Basic Launch testing period. According to the official docs, the testing period ends only when the game has been live for 7 days and played 500 times; if it never reaches 500 plays, it ends automatically after 21 days. Ads are off during this period, so there's no ad revenue.

The official side looks at three metrics; below are the levels commonly seen in well-performing games in the official guide, not a public passing line:

Average playtime above 10 minutes, corresponding to the 90-second short rounds plus multiple upgrade rounds in the design;

Day-2 retention of 10% to 15%, corresponding to auto-save and daily quests;

More than 80% of players who play for a full minute after starting, with a recommendation of loading in under 10 seconds and a package size under 20MB.

Only when the metrics are better than other games on the platform does it enter Full Launch, integrate the SDK, and get ad revenue share. Another path is selling licensing to other platforms, but their exclusivity clauses may conflict with each other, so read the contract carefully before signing.

If you want to publish on WeChat Mini Games

Recently many people in Chinese dev circles have been spreading "WeChat Mini Games single-player titles can be published without filing." I checked WeChat's official docs, and the wording hasn't changed: after the version review passes, you must complete the mini-program filing before publishing. The actual situation is this:

For individuals making free single-player mini games that earn money only from ads (IAA), no game license number (ISBN) is required;

But software copyright, the game self-review and self-inspection report, mini-game filing, and mini-program filing—none of them can be skipped, and the administration bureau review takes 10 to 20 working days;

"Single-player" means it can be played offline. The Three.js in this code is loaded from a CDN; to publish on WeChat you first have to bundle it into the project, then change it to the mini-game project structure;

Some developers in the community report that after signing the single-player commitment letter and passing review, they could click publish before the filing was complete. But the official side hasn't issued any announcement, and if the filing ultimately doesn't pass, it will still be taken down.

So don't go for it because of "no filing required"; prepare the materials according to the process, and polish the game during the filing period—not a bit of time is wasted.

This cat house-wrecking game is still local on my machine and hasn't been submitted. Honestly, even if the model pushes development costs down to a few dollars, most mini games are still hard to hit the mark on the first try. But when the bill for a prototype drops from over ten dollars to a few dollars, the cycle of prototype—look at the data—revise again finally becomes affordable for ordinary creators.

If you want to try it yourself, claim the new-user trial credits here: https://fumolab.ai/en?utm_source=9dta63zy4y

Source

Ayi AI Notes