AI researcher Andrej Karpathy described an experiment in which he asked Opus 5 to transform the opening paragraph of *The Lord of the Rings* into a Three.js rendering, giving the model a one-million-token budget that he estimated cost about $10.
According to Karpathy’s post on X, the model worked for roughly two hours and produced about 5,500 lines of code. The result procedurally rendered elements of the story, placing polygonal assets in three-dimensional coordinates and adding animation. Karpathy characterized the output as imperfect but entertaining, emphasizing that it functioned at all despite the number of spatial and coding decisions involved.
The experiment was presented as a step beyond compact prompts often used to test image or vector generation. Karpathy contrasted it with asking a model to create an SVG of a pelican riding a bicycle, then proposed a longer, less constrained task that required sustained implementation. The resulting program was not offered as a formal benchmark, and the post supplied no comparison run against another model. It is best understood as an individual demonstration.
Karpathy’s larger point concerned the economics of highly specific digital content. A person might not spend hours hand-coding a world tailored to a single idea, he argued, but a model can apply sustained effort at low marginal cost. That could turn projects once dismissed as too custom into reasonable experiments. He imagined generated environments into which players might enter as spectators, non-player characters or figures within a story, comparing the idea to an ephemeral game world created on demand.
At the same time, the demonstration exposed a limitation. Karpathy said the model could not efficiently inspect video or play through the world it had built. To audit the result, it slowly captured screenshots at different points, and that process failed to catch every problem. He linked the visible roughness to gaps in multimodal perception and gameplay capability rather than to an absence of coding stamina.
That qualification is important because the post’s significance lies in both sides of the result. The model maintained a large, coordinated coding task long enough to create a working animated scene, yet lacked a fast feedback loop for experiencing its own product. In interactive development, generating assets and logic is only part of the job; observing behavior and correcting defects can be equally important.
Karpathy’s experiment therefore does not establish a general performance level for Opus 5. It offers a concrete example of extended model-driven creation and a hypothesis about where it may lead: cheap, temporary, deeply customized worlds, provided future systems become better at seeing and testing what they make.
