The AI That Makes 3D Games from Text (and What to Expect)

How a sentence becomes a 3D game

When you flip the 2D/3D toggle to 3D on Arcade Sandbox, two things change under the hood:

1. A real-time 3D engine is injected into your game — a lightweight WebGL renderer with cameras, lighting, and simple shapes (boxes, spheres, cylinders, planes) that runs in any browser with no plugins.
2. Your build runs on Claude's most capable frontier model. Spatial reasoning — where the camera sits, how the world is laid out, what "behind you" means — is the hard part of 3D, so 3D builds get Fable 5, while 2D builds use Claude Opus (more on the model routing).

The model writes the whole game against that engine: the world, movement, camera, enemies, scoring, win state. About two minutes later you're flying, driving, or dodging in a browser tab.

The honest part: what the 3D looks like

This is low-poly, geometric, arcade-style 3D — glowing shapes, stylized worlds, readable silhouettes. Think early-PlayStation-meets-neon, not Unreal Engine. That's not a limitation to apologize for; it's the aesthetic that one-file browser games do brilliantly. Prompt for "photorealistic open-world RPG" and you'll be disappointed; prompt for "a neon wireframe canyon racer" and you'll be grinning. (For the deeper honest ceiling, see 3D game maker online.)

Genres that shine in text-to-3D

• Endless runners & dodgers — corridors of obstacles, satisfying speed ("dodge asteroids in a crystal cave, speed ramps up forever").
• Racers — simple tracks, drift, checkpoints ("hovercar time trials through a floating city").
• Flight & space — six degrees of silly ("defend the mothership from waves of geometric fighters").
• First-person target games — shooting galleries, wave defense ("hold the neon fort against advancing cubes").
• 3D platformers, kept simple — floating islands, collect-and-jump ("hop across sky ruins collecting orbs before the storm").

Prompting in three dimensions

Three details do most of the work in a 3D prompt — say them explicitly:

• Camera: first-person, chase-cam behind the ship, or top-down? The single biggest "feel" decision.
• The verb: what the player does every second — dodge, aim, drift, hop.
• The world's shape: a corridor, an arena, a track, floating islands. Simple geometry = confident level design.

"A 3D endless runner, chase camera, where I'm a paper airplane gliding through an office: dodge staplers and coffee mugs, ride air vents for height, collect paper clips. Speed slowly ramps forever; crashing folds me up with a funny message."

"A first-person wave-defense game on a neon grid: geometric enemies march toward my tower from all sides, I aim and shoot with the mouse, and every 5th wave is a huge slow boss. Synthwave palette, screen shake on hits."

Revising a 3D game

Same as 2D — plain words: "pull the camera back and higher," "make the fog closer and spookier," "double the drift." The engine stays put; the AI patches the game around it. Revisions cost 3 tokens.