Claude Opus 5 builds a 3D scene of The Lord of the Rings in about two hours

What's new

Andrej Karpathy has shown what Claude Opus 5 can do when given a big, open-ended task. He asked the model to turn the opening passage of The Lord of the Rings into a 3D scene that runs directly in a browser. According to Gigazine, the model wrote roughly 5,500 lines of code and took about two hours.

Karpathy shared a 93-second video on X on August 2, 2026. By the time Explainx.ai wrote about it, it had around 2.8 million views.

What's clever

The scene is built with the graphics library Three.js, which is used for 3D in the browser. The model placed characters and objects made of polygons (small surfaces that together form 3D objects) in the space and animated them.

The model also checked its own work by taking screenshots of the result. It can't watch video effectively, so it couldn't review a recording of the scene itself. The audio wasn't made by the model: Karpathy handled it himself using ElevenLabs.

Karpathy points out that tasks like this are changing. According to Explainx, he says language models have all the stamina and patience in the world. So the thinking shifts from "nobody would do this" to "sure, it can just be done, because it's nearly free."

Cheaper or better?

The budget was up to 1 million tokens (chunks of text that the model reads and writes), which according to Gigazine equals roughly $10. The price of Claude Opus 5 per million tokens and subscription prices are not stated in the sources, and there is no direct comparison with competing models.

According to Gigazine, Karpathy believes we are about to move past the well-known pelican test, in which a model has to draw a pelican riding a bicycle as an SVG (a simple vector drawing file). The test was developed by Simon Willison. Gigazine judges the quality to be surprisingly high for two hours of AI work, but says it can't compete with handmade work.

According to Gigazine and The Decoder, Karpathy has put the source code online so the scene can be played in the browser and copied, whereas Explainx's original version stated that only the video had been published.

What it's good for

The demo shows that an AI model can handle a larger, composite project from idea to working result: code, 3D construction, animation and self-checking via screenshots. That could be relevant for people who want to build prototypes, small visual experiences or experiments quickly and cheaply.

It's less obviously suited to finished, handcrafted quality work where details and polish matter. Audio and video still have to be handled by humans or other tools. The sources don't say when, or at what price, regular users in Denmark and the EU can use the model.