Grok 4.5 just released and what better idea than using the free tier to impress a kid.
No? [bad idea, tbh. Specially a single prompt, with no iteration]
To really see how modern agentic runtimes and code generators perform, you need to push them into spatial reasoning and 3D rendering.
is what I read somewhere so I decided to do so.
I asked her what would you like?
A solar system, a dino land or your favourite character in 3D?
She said: A candy land
So, I put Grok 4.5 and Sonnet 5 head-to-head.
The challenge as you know by now was building an interactive, 3D “Candy Land” playground in the browser where users can pick sweet-themed objects and plant them on a rendered field.
Here is a breakdown of how both models handled the prompt, the visual fidelity of their outputs, and the token economics of running the generation.
The Sonnet 5 Experience: Clean Architecture
Sonnet has a reputation for precise logic and structured code. True to form, Sonnet 5 generated a very clean, functional UI code.
The application it built includes a clear header, intuitive action buttons, and a flawless event listener setup for clicking and planting candies on the canvas.
The orbital controls for zooming and dragging work exactly as requested.
However, the visual aesthetics of the 3D objects lean heavily toward basic geometric shapes and flat shading.
It works perfectly, but it lacks visual magic.
The Grok 4.5 Experience: Not so clean code, but more realism
Where Sonnet 5 focused heavily on UI layout, Grok 4.5 poured its parameters into the visual rendering logic.
The biggest differentiator here is the realism.
and this is what the kid liked too.
Grok 4.5 generated significantly better programmatic 3D meshes.
The trees and flowers in the Grok iteration look much more organic and realistic compared to Sonnet’s rigid shapes.
Grok seems to have a better innate understanding of how to use Three.js materials and lighting to create depth, making the playground actually feel like a vibrant Candy Land rather than a wireframe prototype.
Token Economics and Cost
Both models generated entirely correct HTML files packed with custom JavaScript and 3D logic.
Since the input/output tokens consumed were slightly less for Grok 4.5 (~700,~2100) , total generation cost would come around $0.039 compared to Sonnet’s $0.043
Anyway, you can check out the deployed results for Sonnet 5 and Grok 4.5 in my AI Model Lab repository.
Note: The kid found both pretty below average and wanted another iteration xD
They had mentioned if you are generating front-end code where the visual output is the core product, Grok 4.5 takes the win, and I’d like to believe so.
Over the past, I’ve tried building an internal tool/rock-solid UI logic, and Sonnet 5 remains incredibly reliable.
For projects requiring high-fidelity web rendering out of the box, maybe Grok 4.5 justifies the slightly higher input token cost too.
Anyway, this was all just for fun. Might do again, idk.
In case we are meeting for the first time, come over here, it’ll be worth the roller coaster of articles that are gonna come up in the next few weeks.
Have you read?