Karpathy spent $10 and got Claude Opus 5 to build a Lord of the Rings 3D world

Ten bucks and a single paragraph from Tolkien. That’s what Andrej Karpathy fed Claude Opus 5, and what he got back was 5,500 lines of Three.js code—a fully rendered, animated 3D browser scene with audio from Eleven Labs. The model worked autonomously for about two hours, placing objects and animating them without step-by-step human direction. Karpathy posted the source code publicly, so anyone can tinker with it.

The stunt is more than a party trick. Karpathy argues that the old guard of AI benchmarks—like Simon Willison’s famous pelican-on-a-bicycle SVG test or Microsoft’s 2023 unicorn-in-TikZ challenge—are nearly maxed out. Those tests were designed to prove a model could reason spatially about something it had never seen in training data. Today’s multimodal models breeze through them. What Karpathy is probing now isn’t raw intelligence but something fuzzier: a vibe check. Can the model build a coherent, interactive world on a shoestring budget? The answer, at least for a Tolkien forest, is a qualified yes.

But the experiment also exposes a hard technical ceiling. Opus 5 can’t watch its own video output. It had to debug by taking screenshots, which introduced errors—objects slightly misaligned, animations that don’t quite land. That limitation hasn’t stopped other users from pushing the model further, generating first-person shooters and Minecraft clones entirely in the browser. None of this yields a clean score on a leaderboard. That’s the point. Karpathy sees a future where game worlds are spun up on demand for pennies, and you get dropped in as a side character or the protagonist. It’s a vision that sounds like science fiction and costs less than a sandwich.

What’s striking is the economics. A million tokens, two hours of GPU time, ten dollars. That’s the cost of a fully interactive 3D scene that would have taken a human developer days to build. Karpathy, who co-founded OpenAI, led AI at Tesla, and now works on pre-training research at Anthropic, has always been a bellwether for where the field is heading. If he’s spending his time on vibe checks instead of benchmarks, it’s a signal that the old ways of measuring AI progress are starting to feel beside the point.