Exploring how far one person can push AI in game development. ๐ค๐ฎ
AI workflows โข Game dev โข Experiments
Building real games & sharing what actually works. โ๏ธ
1/5 ๐งต
You can make Codex use your own hardware to get WAY more work done without burning through your Codex usage as fast.
And no, this doesn't mean running Codex itself on your GPU.
Here's how ๐
Love it ๐ Honestly it's surprisingly hard to find and connect with other AI game devs on X.
I'm working on multiple games + a few game-dev tools right now. One game is already shipped on Google Play, with several more in the pipeline.
I post progress, experiments and what I learn along the way pretty regularly here.
Definitely following your projects too โ curious to see where you take them ๐ค
Where are the AI game devs on X? ๐ฎ๐ค
Vibe coding a game? Experimenting with AI art, sprites, animation, VFX or agents?
Drop what you're building below ๐
Let's connect, support each other, share workflows and actually ship some games. ๐ค
@thsottiaux@artlevy 70% off for new partners ๐
Meanwhile us loyal Codex users burning through our weekly limits directly in ChatGPT:
Where's our discount, Tibo? ๐
Benchmarks don't tell the whole story.
I've tested Grok 4.6 vs GPT-5.6 Sol on actual game dev. Grok is fantastic for UI, but Sol is still ahead for me on sprite sheets, animations, asset implementation and VFX.
Here's one test: same game, same task โ build a AAA-quality nuke ๐
Grok went mostly code/VFX. Codex could use sprite sheets + code and implement the whole thing directly.
If xAI fixes Grok's sprite sheet generation + transparency/cutting/implementation workflow, THEN this gets really interesting ๐
https://t.co/aG5lMyjAPZ
โ๏ธ CODEX 5.6 SOL vs GROK 4.6 โ NUKE TEST โข๏ธ
I gave both models the same challenge for my zombie game: create an AAA-quality nuclear explosion, with complete freedom over how to build it.
After multiple iterations, this is where they ended up:
๐ต Codex 5.6 Sol โ generated its own sprite sheets and ended up using a hybrid approach: sprites + code/VFX.
โซ Grok 4.6 โ went almost entirely with code-based VFX, since it still struggles with generating clean, usable sprite sheets.
Both were pushed through multiple iterations, and IMO both still need improvement. This was actually a surprisingly hard exercise for both models.
But I'm curious:
Which nuke do you prefer โ Codex or Grok? ๐
Videos below ๐
I've tested both extensively for game dev and GPT-5.6 Sol is clearly ahead for my workflow.
Grok is excellent for UI โ I actually prefer it there.
But for sprite sheets, animation, asset implementation, VFX and working across an actual game project, Sol is on another level for me.
Benchmarks don't capture that.
@PussyBlaster10k@thsottiaux Yes, 100%. That's actually one of the reasons I prefer it for game dev.
Codex gives me real transparent PNGs, then handles the cutting/alignment and implementation too. No checkerboard pretending to be transparency ๐
Gemini/Nano Banana can make good images, but Gemini coding has been pretty bad for my game-dev workflow.
And transparency drives me crazy ๐ Ask for a transparent PNG and sometimes it literally generates the checkerboard squares INTO the image instead of actual transparency.
Then you still need other tools/models to clean, align and implement the sprite sheet.
Codex + GPT-5.6 Sol does the whole loop for me: generate โ cut โ align โ animate โ implement โ fix.
Sol 5.6 is definitely the one I'd try.
@Auditormadness9@thsottiaux That's my issue with Claude for game dev โ you still need a third-party image generator for the assets.
Then you're jumping between tools for generation โ sprite sheets โ cleanup โ implementation.
Codex keeping that workflow in one place is what sold me.
@DannyLimanseta which agent you used for this game and how was the token consumption in average? Is it more interesting to do things manually cost-wise? Thanks
Why H3 specifically though? ๐
Isn't the real unlock the workflow itself?
Video โ extract key poses โ remove/key background โ pack into sprite atlas โ state machine.
In theory the source could be ANY decent AI video model (or even an existing animation), as long as the character stays consistent enough.
Would be interesting to test this across different video models.
5/5
So if you already have a decent GPU, you can basically build a little AI team:
Codex = boss ๐ง
Local model = worker ๐จ
Your GPU = pays the worker ๐
1/5 ๐งต
You can make Codex use your own hardware to get WAY more work done without burning through your Codex usage as fast.
And no, this doesn't mean running Codex itself on your GPU.
Here's how ๐
4/5
The important part:
Your GPU does NOT magically give Codex more tokens or increase your official usage limit.
You're simply moving some of the workload to a model running locally on your own hardware.
Less work for Codex = your usage can go much further.