I tried GPT-6 Astra and while the general vibe feels the same as GPT-5.6 Sol if you don’t give it harder prompts, one of the things I saw with GPT-6 demos are actually managed to crack blender renderings and very very detailed voxels.

I am shocked with this:

this is astra btw.

first model to visibly leave me at a loss for words https://t.co/pC2nmFcjQp pic.twitter.com/eZ7vjWICTy

— adi (@adonis_singh) September 4, 2026

Therefore since I have a subscription and 3 banked resets to spare, I had decided to push this model to something I’d never done with LLMs before.

MS Paint

#

I want to see it for myself, so I decided to ask Codex to draw something in https://jspaint.app

Fast, practical, it tests spatial and geometric reasoning including its speed. Before CUA models would 100% stumble with this, not even Gemini… and SVG generation had already been a benchmark in the past so if SVG was a struggle then there’s no way these models can draw using first principles right?

But then as I tried Astra myself to draw something in MS Paint, my jaw was dropped and I was laughing with excitement, so here are the things I asked GPT-6 to draw

These are ran through medium reasoning effort.

Scott howl

#

Monster prom is certainly a niche game, but this is a good way to test model’s world knowledge and generalization when it comes to connecting concepts such as the character and the shapes, color, and attributes needed to draw.

For reference, this is Scott Howl:

And this is what GPT-6 cooked

Jaw dropped it’s already legible and recognizable, this is something would have considered impossible a year ago to achieve this, and it’s not even a GPT Pro-level model! And it only took 6 mins too, impressive for its speed for doing CUA drawing (which is historically slowest task an LLM would do).

It got the fine details right too including his facial hair.

Twilight Sparkle

#

Last year and two, a model tries to even render a unicorn SVG:

It is improving in late 2025:

And yes we get to the point these models can render SVG Xbox or PS5 controllers as if its ai generated image.

But CUA is a different thing, it’s slow and challenges vision and spatial reasoning.

But if we had a pretty emerging SVG horses in 2025, then Astra already leapfrogged drawing a pony using CUA Paint, i’ve rarely seen one model where it even tried to draw it (and would be costly and takes probably days to complete). And it took 6 mins too.

Yeah… it can already draw a horse

Complete with stripes, horn, and ponytail though it didnt include her wings since she was an Alicorn in later seasons.

Mordecai and Rigby

#

Yeah, not just one characters but two characters, and it drew quite accurately, including their stripes. It also took almost the same time as other character drawings.

LibreOffice Calc

#

Yep, I did also try making Excel pixel art, I asked for “Scott Howl” character and it turns out it’s not looking bad.

This is without any references attached.

Because of the lack of support for Computer use in Linux version of ChatGPT, the hacky workaround is to setup NoVNC server with TWM and TigerVNC then use a built-in browser to control it. I could have used Excel online but I don’t have time to login with my MSA.

Usage

#

For plus plan, and with medium reasoning. 3 CUA drawings already burned 100% of my 5 hour usage limits.

So yeah worth buying pro or wait for reset.

Conclusion

#

This is the part where I’d say we’re near AGI one small step at a time.