Fine But Why
Newsletter

Benchmark

Four games in /goal mode

Muse Code with Wonder, one prompt per game, hands off. Every prompt ends the same way: keep refining until it looks beautiful.

Yair · Oct 13, 2026

The lawn. 10m46s, about $2.58. The middle is sped up 20 times.

In /goal mode Muse Code keeps working on its own until it decides the goal is met. I gave it one prompt and didn't help. Every game is one index.html with three.js from a CDN, and every prompt ends with:

don't stop at the first version that works: play it, find what looks or feels off, and keep refining until it looks beautiful

The player in each clip is a script pressing the game's own keys. Muse Code has no cost meter, so the cost is the tokens in its session log at list price. Model muse-wonder-1.0 at xhigh, with threejs-skills in the project.

Mow the lawn

10m46s, about $2.58, 38 model calls, Muse Code 1.4.6. That's the clip at the top.

build a satisfying 3d lawn mowing game of my backyard: a big overgrown lawn of thick grass blades swaying in the wind, a wooden fence around it, a couple of trees, flower beds and a small shed. i push a lawn mower with wasd and the camera follows behind it. wherever the mower goes the grass gets cut short and leaves light and dark mowing stripes, with clippings spraying out behind it. show how much of the lawn is mowed and celebrate when it is done.

68,000 blades of grass sway in the wind. The mower leaves light and dark stripes and sprays clippings out the back, and there is confetti when the lawn is done. On X.

Wash the patio

10m16s, about $3.85, 46 model calls, Muse Code 1.4.6. The middle is sped up 6 times.

build a satisfying 3d pressure washing game of my back patio: big stone patio tiles caked in dark grime, moss in the cracks and old leaves, a wooden deck, garden chairs and a table, a grill, potted plants and a brick wall with ivy. first person, i walk with wasd, aim with the mouse and hold the left mouse button to spray. a jet of water blasts the grime off wherever it hits and reveals clean bright stone underneath with a crisp edge, with spray mist, droplets and wet surfaces that glisten. show how clean the patio is and celebrate when it is done.

New this time: a small art-direction skill next to the three.js ones. It says to pick a look first, then screenshot the game and fix the worst thing, at least four rounds. It did four: grime that washed out, spray that blew out the picture, a wet layer that looked like ice, and a lance that was far too big. On X.

Fix the station

22m37s, about $5.24, Muse Code 1.4.7. Real time, nothing sped up.

build a breathtaking 3d spacewalk game: i'm an astronaut outside a space station in orbit, with the earth huge and glowing below me, a thin blue line of atmosphere on the horizon, the sun rising over its edge and a sky full of stars. first person from inside my helmet with a faint visor reflection and a small hud. i float with wasd, go up and down with space and shift, look with the mouse, and a tether clipped to the station sags and swings behind me. three solar panels on the station are broken and sparking: fly to each one and hold the left mouse button to weld it, with bright sparks, and the panel lights up blue when it is fixed. show how many panels are repaired and celebrate when all three are done.

Same two skills as the patio. Its screenshot rounds fixed the station being out of frame, a blown-out hull, the tether hanging in view, a flat earth and panels that were too dark. I added one CSS hint so the visor overlay got its own layer while recording. It looks the same and keeps the clip at 60 fps. On X.

Snakes in the grass

32m39s, about $6.48, Muse Code 1.4.7. The middle is sped up 25 times.

A reply on X suggested snakes in the lawn. So: the same backyard, that line word for word, plus snakes in the grass and bugs buzzing low. Steer around them, a hit costs a life, three hits and it's game over. And the last line: keep refining until the snakes and bugs look great and move like real animals.

It looked at its own screenshots about fifteen times. The snakes vanished in the tall grass, so it gave them bold bands, a head that rises when they stop, and little puffs of grass where they move. The bugs got bigger. The first try ran without a browser to look at its work, about $2.18, so I ran it again with one. The script in this clip reads where the snakes and bugs are and steers around them with the game's own keys. On X.

All benchmarks