What changed in the stack

In conventional 3D development, assets are assembled from libraries, textures are loaded as image files and physics is handed to an engine. In the demonstrations shared over the past few days that sequence disappears: geometry is built by code that computes points and surfaces at run time, textures exist as shaders and therefore as more code, and the physics and controls belong to the same generation. Because it all runs on Three.js in the browser, the output fits in a single HTML file.[1]

For a developer the difference has a concrete payoff. Instead of an asset pipeline there is a file that can be opened and changed; Matt Shumer was able to publish both the prompt and the source of his shooter. Where the cost accumulates also becomes visible: Pankaj Kumar wrote that his Minecraft clone, with an infinite procedural world, fifteen biomes and two game modes, consumed 25 million tokens. The expense moves from asset licences and engine setup to a generation budget.[1]

What the comparisons did not hold constant

Every cross-model comparison here is a single run scored by the person who ran it. Chetaslua put the same prompt through Opus 5, GPT-5.6 Sol and Kimi K3 and said the others returned fewer working parts; atomic.chat had four models build three physics scenes and reported that only one handled all three convincingly; ChrisGPT repeated a similar request a year apart. Anthropic's line about "much stronger visual outputs" is a vendor statement. In none of these was the prompt set, the number of runs or the scoring criterion held fixed.[1]

That does not make the demonstrations worthless; it makes it impossible to separate how much comes from the model and how much from the choice of prompt and the browser target. The setup that would separate them is not complicated: a shared prompt set, several runs per model, a scoring rubric written in advance and failed attempts reported alongside the successes. Until such a comparison is published, what we have is individual files that were seen to run.[1]