Gemini 4 Argon vs Muse Wonder in game development. Same prompt, two completely different games.

@rehan_shei bro we know u just made the bottom one & then video-rendered the top one 😂

@rehan_shei you can play the game here:

@rehan_shei How did you keep the choreography the same

@rehan_shei We used the two models to create the games and then used an agent to play them. Also the games share the same logic, boss attacks etc. to make them comparable!

Surprised?

You can play the game made by Gemini 4 Argon here:

This seems a lot like a video generated using the low quality as baseline.

well, you can test the game yourself:

holy mog

Fucking lieing I used argon and it couldn’t even make a simple 2d game it was ass and looked like the designs were drawn by a 2 year old

Prompt?

There is no way they got the same game from that simple prompt. This looks like clear interations on the same game that was previously. Boss has the same exact attacks, the only difference is the visuals.

AI game demos always hide the unplayable mess underneath

brooo I keep getting so close

Anyone who ever used Gemini and antigravity would say this is a lie, Gemini can barely get code correct, all of my agents double check Geminis work which is wrong more often then not and if something failed he fakes output and tries to cover it up. Maybe one prompt after you create assets and do all the rigging and animations hand holding the AI. Prove me wrong what did you start with and what prompts did you use?

How did the mechanics turn out exactly the same? It’s literally the same game, only difference is the assets. Find it hard to believe only prompts were fed to each model there had to be a defined baseline here

@demishassabis Yea b/c Muse Wonder wasn’t trained on 3D animations, this comparison is useless

These comparison are suspect, potentially fraudulent, because in everyone of them the character's position, animation, and rotation in both version is exactly the same. That's extremely hard to do manually, and it's much harder to do consistently. So I suspect these comparisons are fraudulent.

no way, the animations are identical, this is cap

The Gemini version makes me want to play games more. The Muse version makes me happy I’m not playing games as often.

How's it go against opus 5.5?

comparison coming soon!

Super keen to see it! Vibe coding the best Pokemon game ever made for like 1 year full time / 2 years now

Sharing the boss logic keeps this focused on presentation. A model can dress up the same fight beautifully without having designed the fight itself.

same prompt giving two totally different games is the real test of a model, since it shows how much it plans structure versus just copying a template. curious how each one handled the game loop and not just the visuals

Where is the prompt? Why not to share prompt by default when you are comparing models? How everyone can be sure that you are not playing with us? :)

it's impossible to beat the boss in your demo, at some point his health bar just stops going down no matter how much you hit him.

@pigeon__s There’s no way two different models could produce such similar game mechanics on a frame by frame level

I'd like to try an Astra run at Ultra effort. Could you share the exact prompt, starting files and agent setup? Your note that both games share boss logic makes the starting point especially useful for a fair comparison.

nice, now do opus 5.5

how is it so similar though lol

Bro Muse can't even make it to PS1 levels 😂

camera movement is quite messed up but everything else looks impressive, and extra impressive for a one prompt wonder

Argon looks amazing

Can you share the prompt?

@grok can we make games with Gemini argon and is it paid

Just played through the game link bro shared, and it's actually real and playable Mind-blowing bro! 🤯

同提示词对比最大的变量恰恰是提示词本身：同一句话在两家模型的「舒适区」里位置不一样，赢的往往是更吃这句提示词的那家。真想分高下，得固定3-5个不同风格的提示词各跑3次，看一次成功率和方差。单发demo看看就好，别当购买指南。

Same prompt, two completely different games is the perfect stress test. I ran a similar bake-off last month and the models diverged so hard I thought I'd messed up the prompts. Turns out that's just how different their priors are.

Same prompts producing different games is a useful test of model bias. Did you compare the code and game loops too, or only the final visuals?