Forget the one-shot pelican. AI's new benchmark is a 2-hour Middle-earth build.

The public bar for what counts as a "frontier AI" has just visibly moved. Andrej Karpathy has a new project: read the opening of The Lord of the Rings, spend about two hours and roughly a million tokens writing 5,500 lines of Three.js code, and produce a Middle-earth you can actually walk around in a browser.

Three.js is the open-source toolkit that turns code into 3D scenes, the same plumbing behind a lot of web games and product configurators. Karpathy's project page posts the work as a public "vibe test": prompt the model with the first chapter of Tolkien plus a Three.js instruction sheet, then judge the run by whether the world actually loads, animates, and stays coherent over time.

The run consumed about 1,000,000 tokens, the rough unit of text an AI reads and writes, and roughly 5,500 lines of generated code. The Decoder's independent English report pegs the API cost at about $10. Two hours, a million tokens, a 3D world. The old test, Simon Willison's single-shot SVG of a pelican on a bicycle, used to fit on one screen and took seconds.

The pelican was a one-paragraph prompt with no persistence: a model either drew the bird and the bike in a sensible arrangement or it did not. The Middle-earth run has to keep state across hours of generation. Lighting, camera path, terrain, character placement, and the way the scene continues to look coherent as new code lands on top of old code, those are now the things the community is publicly measuring.

Microsoft's 2023 "Sparks of Artificial General Intelligence" paper used a GPT-4 unicorn drawn in TikZ, the LaTeX graphics language, as evidence the model could reason about spatial relations. A 2D vector of one horned animal then; a 2-hour, 5,500-line 3D world now.

Chinese tech outlet 量子位 rounded up the parallel community builds that broke in the days after Karpathy posted: a low-poly San Francisco called Earth Online, a New York 3D model attributed to GPT-5.6 Sol and Fable 5, a 14-song Kanye West virtual concert, and an interactive 3D game demo built off Opus 5. None of these are products. They are stress tests.

The model writes the world, but it does not get to live in it. Karpathy's own check is to take periodic screenshots and inspect them; the AI has no continuous video of what it is rendering, no way to "see" whether a character is clipping into a hill or whether a shadow lands in the right place. In a comment thread on the project, he pointed at a workflow where procedural code handles blocking, camera, and layout while a video-to-video model like Seedance handles textures and lighting, a division of labor that papers over the missing eyes.

Audio was not generated by the model either. The final voice track was done by ElevenLabs, the third-party text-to-speech service. The code is open-sourced on Karpathy's site, so anyone with an API key can rerun the same prompt and watch their own model either hold the world together or fall apart.

A pelican on a bicycle can be a clever hack. A 2-hour, 1,000,000-token build of a readable Middle-earth is closer to what "can it actually build something" means as a research question, and it is the kind of test a reader can now hold every next "AI can now…" headline against: did the model just draw a clever picture, or did it keep a world running?