Three models each generate a horse; Kimi's version wins this round

There's a test circulating quite a bit on X. The same prompt asks three models to each generate a separate HTML file that uses three.js to render a horse running on a grassland. No build tools, no external assets, and all physics parameters and animation logic packed into one file.

The three models are Kimi K3.1 Pro, Claude Opus 5.5, and GPT-6 Astra.

In Kimi's version, the limb rhythm and body proportions are closest to normal running. Several people in the comments said Opus 5.5 looked the worst, and some netizens guessed the test used a low thinking setting. Astra isn't necessarily that good either, but the fact that its hind legs run backward gave it another kind of attention in the comments.

One netizen said the Kimi version was "smoother animation, better proportions, less jank." Another put it more directly: open weights at this level are trouble for closed-source models.

A single-file three.js test has one advantage: the model can't hide behind project structure. Whatever the code looks like when it runs is right there. Physics choices, animation parameters, rendering logic—all laid out in one file. Netizen jonathan said in a reply that this kind of test has "nothing to hide behind," and physics choices are exposed immediately.

Later someone upgraded the test. The prompt became: Xbox controller → bicycle → pelican → horse → bicycle → @theo in Rio. A chain of objects stacked together, to see how the model handles occlusion, layers, and positional relationships.

This one is closer to a stress test. Rendering a single object is still fairly controllable, but once the chain gets long, the model's understanding of space and object relationships is dragged out for inspection. A netizen commented that if this chain can really be connected accurately, that would be a manifestation of general reasoning.

However, using a single three.js file as a benchmark also has limitations. It tests code generation and physical intuition in a specific scenario; it doesn't mean the models have opened up a gap on all tasks. The prompt precision, thinking setting, and sampling parameters each model received are also not fully transparent.

In the comments, some said Kimi mogged them, some said Anthropic is in trouble, and others joked that open source is about to be banned. These reactions may not be objective, but at least they show one thing: an open-source model tying or even narrowly beating closed-source frontier models on an intuitive task like visual programming is more shareable than a leaderboard.