Karpathy had Claude Opus 5 generate a 3D scene from The Lord of the Rings as an informal test of capabilities
Andrej Karpathy had Claude Opus 5 generate a complete 3D scene in Three.js from the opening paragraph of The Lord of the Rings — roughly two hours and about 10 dollars produced 5500 lines of code. He describes it as an informal vibe check, not a benchmark.
Andrej Karpathy, co-founder of OpenAI and now a researcher at Anthropic, had Claude Opus 5 turn one paragraph from the opening of The Lord of the Rings into a 3D scene running in a browser. The model worked for roughly two hours, had a budget of one million tokens (approximately 10 dollars), and generated about 5500 lines of code using the Three.js graphics library. Karpathy published the source code online, so others can play the scene and modify it.
The model placed and animated objects in the scene on its own, but made mistakes because it could not watch its own generated visual output and had to rely solely on screenshots. Eleven Labs provided the soundtrack for the demonstration; the rest is the work of the model.
According to Karpathy, the popular test of model capabilities used so far — having them draw an SVG pelican on a bicycle (a test devised by Simon Willison) — has almost exhausted its usefulness as an indicator of progress. Karpathy also recalls an earlier counterpart from 2023, when Microsoft had GPT-4 draw unicorns in TikZ as evidence of spatial reasoning in the Sparks of AGI paper. He therefore describes the scene from The Lord of the Rings as merely an informal “vibe check”, rather than a rigorous benchmark — much like other users who have Claude Opus 5 write entire browser games, from a first-person shooter to Minecraft clones.
Karpathy sees potential in this for inexpensive, custom-generated game worlds in which users could be placed as supporting characters or main protagonists. He previously served as head of AI at Tesla and led the education startup Eureka Labs; since May 2026, he has been researching pre-training at Anthropic.
Why it matters
This shows that a current-generation model can independently design and program a larger interactive 3D application from a short text prompt at low cost and within hours, suggesting a shift in the kinds of tasks that can be entrusted to models without detailed instructions. At the same time, this is only an informal demonstration by an individual, not a measured benchmark or a finished product — errors caused by the model being unable to see its own video output show that limitations remain.
Relevant practical impact
What this means
For individuals
The experiment shows that Claude Opus 5 can turn a short literary text into extensive, functional code (thousands of lines) for a browser-based 3D scene in a few hours and for a few dollars, which can serve as a reference point when estimating what creative programming tasks can be assigned to a model today.
Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.