Claude Opus 5 spent two hours writing 5,500 lines of JavaScript code to procedurally render the opening of the Lord of the Rings as an interactive 3D scene, all within a $10 token budget. The experiment, published by AI researcher Andrej Karpathy on August 2, reveals a workload class that barely existed six months ago: long-horizon agentic tasks where the constraint is a dollar budget rather than a single prompt completion.

The setup was deceptively simple. Karpathy gave Opus 5 the first paragraph of Tolkien's novel, a 1 million token budget (roughly $10), and asked for a three.js render. The model then orchestrated polygon assets in three-dimensional space, wrote animation code, and iterated through corrections across a two-hour run. The result is janky and unfinished, but it demonstrates something genuinely new: an AI system behaving like a contractor with unlimited stamina and a deadline expressed in dollars rather than hours.

What Makes This Different From Previous AI Tests?

For years, AI researchers tested large language models (LLMs) with simple, single-artifact tasks. The classic example was asking a model to generate an SVG image of a pelican on a bicycle. These tests were quick and easy to evaluate, but they measured only whether a model could succeed once, not whether it could sustain complex work over time. Karpathy's experiment flips that entirely.

The developer community discussing the experiment identified four distinct perspectives on what the test reveals:

• Benchmark Design: Many developers argue that single-artifact tests were always a gimmick, and that long-form generation with an explicit budget is a strictly better stress test because it measures sustained execution, cost discipline, and self-correction across multiple iterations.
• Game Development Skepticism: Former game developers pushed back hard on the vision of AI-generated ephemeral worlds, noting that token-generated games lack real mechanics, tuning loops, and the hours of playtesting that make games engaging beyond novelty.
• Entertainment as Shared Experience: Some commenters argued that most people prefer consuming content with others rather than producing it alone, suggesting that AI-generated media will coexist with blockbusters rather than replace them.
• Source Credibility: A minority questioned whether the post was marketing given earlier statements about reliable agents being a decade away, though the majority noted that eight months of progress on capability curves is not the same as being wrong.

Where Opus 5 Actually Struggled

The most revealing part of Karpathy's analysis is not what the model succeeded at, but where it failed. Opus 5 could write complex, coordinated code, but it could not watch its own render in real time. Instead, it had to take screenshots at different points in the scene, slowly and painstakingly, with retries along the way. This created the jank that Karpathy freely admits to in the final artifact.

"Agents are not yet able to efficiently and natively perceive videos or play games within them," Karpathy noted, adding that "raw multimodal self-audit is one of the capabilities that is still genuinely lacking."

Andrej Karpathy, AI Researcher

For any team building long-running agents, this observation is critical. The bottleneck is not code generation or creative problem-solving. It is verification. An AI system can produce artifacts at scale, but auditing its own work remains slow and error-prone.

How to Design Better Long-Horizon AI Tasks

• Wire Explicit Observation Loops: Rather than asking a model to self-inspect from memory, integrate deterministic observation loops like those used in browser automation or 3D rendering. This lets the model validate its output by construction instead of by screenshot.
• Use Rigid Intermediate Representations: CAD-style tool integrations work precisely because geometry is anchored to a deterministic interpretation engine. The more an agent's output space is constrained by executable ground truth, the less it needs to see to know it succeeded.
• Frame Sessions as Budget-Constrained Runs: Treating a session as "a run with a budget" instead of "one prompt" changes how you review it. The artifact is judged on what bounded compute produced, failures are categorized as cost or capability problems, and jank becomes a spec for iteration rather than a disappointment.

Why the Economics Matter

The real significance of Karpathy's experiment is economic, not technical. A bespoke 3D interpretation of one paragraph of a novel would previously have required a studio budget and a production schedule. At a marginal cost of roughly $10, the question stops being whether it is worth building and becomes why you would not.

This shifts the entire calculus for creative and technical work. LLMs have all the patience in the world, so a task that seemed impossible or prohibitively expensive becomes feasible. The constraint is no longer time or human labor. It is the model's ability to verify its own work and iterate toward a usable result.

The token-budget frame is a genuinely useful discipline for evaluating AI systems. It moves beyond single-number benchmarks and instead measures what a bounded amount of compute can actually produce in the real world. For teams building agents, this is the lens that matters most.