Andrej Karpathy shared on X on Saturday that he tested Anthropic’s Claude Opus by giving it the opening paragraph of The Lord of the Rings, a large token budget, and a request to create a Three.js rendering of the story. The model worked autonomously for approximately two hours and produced roughly 5,500 lines of procedural code, Benzinga reported.
“We’re starting to leave the territory where you’d test an LLM by e.g. ‘create an svg of pelican on a bicycle,’” Karpathy wrote. The model had to “place and orchestrate various polygon assets in (x,y,z) coordinates and write code that animates it all.”
What the Test Showed
Karpathy described the output as “kind of janky but fun.” The significance, in his view, was not the visual quality but the scope of autonomous work: sustained multi-step reasoning, spatial planning, asset placement, and animation logic across thousands of lines of code, all without human intervention between steps.
“LLMs have all the stamina and patience in the world,” Karpathy wrote. He argued this enables projects “no one would ever do” because the human time investment would be prohibitive. A person could build a 3D rendered world from a novel passage, but almost nobody would spend the hours required. An AI model with a sufficient token budget will.
The Broader Coding Shift
The test lands in the context of rapidly accelerating AI code generation across the industry. OpenAI President Greg Brockman said in May that AI tools went from writing about 20% of code to 80% within a single month at OpenAI, according to Benzinga. Airbnb reported that AI generated 60% of its new code in Q1 2026, with one engineer handling what previously required a team of 20, per Benzinga.
What Karpathy’s test adds to these production metrics is a different dimension: not speed on routine tasks but sustained creative output. Writing 5,500 lines of procedural 3D code from a literary prompt requires spatial reasoning, artistic interpretation, and multi-file orchestration over an extended execution window. That is closer to what agent builders are targeting with long-running autonomous tasks than the autocomplete-style code generation that defined the first wave of AI coding tools.
From Autocomplete to Autonomous Sessions
The gap between “generate a function” and “build a world from a paragraph” is the gap between copilot and agent. Karpathy’s framing points to a near-term future where the bottleneck for AI-generated software shifts from capability to context window and token budget. If the model can sustain coherent output across thousands of lines over hours, the constraint becomes how much compute you are willing to spend per session, not whether the model can do the work.
For teams building on agent frameworks like OpenClaw, the implication is practical: the infrastructure for multi-hour autonomous coding sessions needs to handle sustained execution, state persistence, and cost monitoring at a level most current deployments are not designed for.