AI researcher Andrej Karpathy, who joined Anthropic earlier this year, tested Claude Opus 5 by giving it the opening paragraph of J.R.R. Tolkien’s The Lord of the Rings and a $10 budget allocation of 1 million tokens, asking it to render the scene as an interactive 3D web environment using Three.js. Over two hours, the AI model autonomously wrote 5,500 lines of JavaScript code, producing a low-poly, procedurally generated 3D rendering of the Shire with hobbit holes, trees, rivers, character assets, and ambient animations synchronised with narrated book excerpts. Karpathy noted the visual rendering had noticeable flaws but said the experiment shows a fundamental shift in testing AI capabilities beyond simple prompts like creating an SVG of a pelican on a bicycle. He highlighted that the model operated with computational patience humans rarely dedicate to hyper-specific tasks, turning "no one would ever do this" into "sure, why not, it's ~free." He expressed excitement about creating hyper custom worlds, such as an ephemeral GTA of X on demand. However, the test also revealed significant gaps: because LLMs lack native, real-time video perception and the ability to seamlessly interact with live environments, the model struggled to audit its own visual output.