These are AI-generated games you can play right now, not trailers. Four frontier models each built three games from a single prompt: a physics toy, a first-person shooter and a top-down MOBA. Nobody touched the code afterwards. Watch them below, then press Open and play.
What games did the AI make?
Three genres, four versions each.
- A 3D first-person shooter. An arena with cover, waves of enemies, recoil, reloading, a HUD and a game-over screen. Claude Opus 5.5 built Iron Tide, a neon city at night; Claude Fable 5.1 built Duskfall, a rooftop under a violet sky; GPT-6 Astra built Dead Signal and GPT-6.1 Sol built Deadlink, two block-built yards that both put a door marked Relay in front of you.
- A one-lane MOBA. A League-style lane seen from above, with a click-to-move hero, four abilities on Q W E R, minion waves, towers and an AI enemy hero. Astra's Emberfall is a low-poly forest lane; Sol's Aetherfall is a dark, sparse arena. Both Claude models named theirs Riftline: Opus 5.5's is a lamp-lit forest lane off a stone plaza, Fable 5.1's a pine forest with a river crossing, and both have a shop.
- A physics toy. Twenty numbered balls bouncing inside a hexagon that turns once every ten seconds, with the collision maths written from scratch.
How good are AI-generated games?
Complete, and uneven in feel. Every game has a working loop, a way to win or lose and a restart. The Claude games are richer to look at, with more lighting, effects and interface, and took over an hour each. The GPT games arrived in 8 to 25 minutes and are plainer. Under our fixed input script, most heroes die at least once: Claude Opus 5.5's shooter player falls at 25 seconds and GPT-6.1 Sol's at 35, and in the MOBA the scripted push into an enemy tower costs Claude Fable 5.1's hero and both GPT heroes a life, while Claude Opus 5.5's hero survives.
How were the games made and recorded?
With the same rules for every model. Each received the identical prompt, ran once in its vendor's coding agent with a 150-minute limit, and had to deliver files that run from disk with three.js provided as a local copy and no other assets. We record games frame by frame on a virtual clock at 60 frames per second, with randomness seeded identically, so a slow recording machine cannot make one game look choppier than another. The methodology has the details, and the MOBA and shooter briefs have the full prompts.
Which AI should make your game?
A Claude model if you want the most complete-feeling prototype and can wait over an hour; a GPT-6 model if you want something playable in twenty minutes to iterate on. For the Claude-only and ChatGPT-only collections, see Claude games and ChatGPT games.
Four MOBAs, names hidden. Play each for a minute, pick one, then reveal.
Take the blind test
Frequently asked questions
Are these AI-generated games actually playable?
Yes. Every game opens in your browser in a sandbox, exactly as the model wrote it, and responds to your keyboard and mouse. The videos are recordings of the same files under a fixed input script.
Which AI made the best game?
The Claude models built the most atmospheric games, with richer lighting and effects, and took the longest. The GPT-6 models built plainer games much faster. Play them or take the blind test to decide which you prefer.
Did AI generate the art and sound too?
Yes, in code. The briefs allowed no asset files, so every model, texture and sound effect was generated procedurally: geometry built in three.js, textures drawn on a canvas and sound synthesised with the Web Audio API.
How long does it take AI to make a game?
From 8 minutes for the physics toy to the full 150-minute limit for the most ambitious 3D shooter. Most games took between 13 and 75 minutes.