One prompt, 690 million tokens and $423 produced a playable 3D racing game

One prompt, 690 million tokens, $423, and a playable 3D racing game came out the other end.

The one-shot build

Developer Vyom says he gave Opus 5 a single prompt of roughly 2,000 words, written more like a product requirements document than an instruction, and let the model run in one shot. His own summary: “690 million tokens, $423 and just 1 prompt is all it took.”

The output is a cel-shaded arcade speedboat racer built with Vite, TypeScript and Three.js. The meshes, textures, water, audio and UI were generated in code rather than downloaded as assets, which makes the token count less absurd than it first sounds: the model was not wiring up a game around existing art, it was producing everything.

The $5 counterpoint

Then the internet did what it does. Another developer, Anul Agarwal, rebuilt a comparable version in Codex on GPT-5.6 using two prompts and about five hours, and put the cost at roughly $5. That figure deserves its asterisk: it is his own estimate converted from subscription usage rather than a metered API bill, so the 423-against-5 comparison is directional, not a benchmark. The remake also reads rougher on water detail and character movement.

The price of not touching the keyboard

The interesting number is not either price tag. It is the structure of the difference. The expensive run was one prompt with zero human corrections. The cheap run was a person steering twice and staying in the loop for five hours. Read that way, the two builds are measuring different products: one is the cost of a game, the other is the cost of full autonomy.

On this single, loosely controlled data point, the tax on autonomy runs roughly two orders of magnitude. That ratio, not the $423, is the figure worth tracking as models take on longer unsupervised runs.