Karpathy uses Claude to generate 3D game world for $10
Anthropic researcher Andrej Karpathy used Claude Opus 5 to generate a 3D "Lord of the Rings" scene for ten dollars, demonstrating the potential for cheap, on-demand game world generation.

Andrej Karpathy, who co-founded OpenAI, previously led AI at Tesla, founded Eureka Labs, and joined Anthropic in May 2026, recently tested the capabilities of the Claude Opus 5 model. He tasked the system with translating a single paragraph of "The Lord of the Rings" into an interactive 3D browser scene. Over a two-hour period, the model generated 5,500 lines of Three.js code to render the environment, utilizing a budget of approximately one million tokens, which cost roughly ten dollars. The resulting scene featured automated object placement and animations, with audio provided by Eleven Labs. Karpathy has shared the source code online for others to inspect and modify.
While the experiment yielded an impressive result, it was not without flaws. Because Claude Opus 5 cannot directly view its own video output, it had to evaluate its work using static screenshots, leading to rendering errors. Karpathy described the experiment as a "vibe check" rather than a rigorous benchmark. He noted that older, text-based tests—such as Microsoft's 2023 experiment prompting GPT-4 to draw a unicorn in TikZ, or Simon Willison's pelican test requiring an SVG of a pelican on a bicycle—are becoming obsolete as multimodal models max out these challenges.
For AI practitioners and game developers, this experiment highlights a shift toward highly accessible, automated world-building. While professional developers would not typically write code this way, the extremely low cost of ten dollars makes on-demand, personalized game environments a viable reality. Users could soon generate custom virtual worlds on the fly and enter them as active participants. This capability is already being explored by others using Opus 5 to build complete browser games, ranging from first-person shooters to Minecraft clones, signaling a new era of rapid, LLM-driven software prototyping.
This is our own summary of reporting by The Decoder



