Skip to content
GlanceHall
  • Builds
  • Models
  • Compare
  • How we test
Get the next round
  • Builds
  • Models
  • Compare
  • How we test
  • Journal
  • About
Get the next round

We don't rate. We let you see.

contact@glancehall.com

Watch

  • All builds
  • Models
  • Head to head
  • Journal

Trust

  • How we test
  • About
  • FAQ
  • Contact

Legal

  • Privacy
  • Terms
  • Cookies
© 2026 GlanceHallIndependent. Not affiliated with Anthropic, OpenAI or any model vendor.
Journal/Can AI build it?

Vibe coding games: three games, four models, one prompt each

Is it possible to vibe code a game? We gave four AI models three game prompts, one attempt each, no edits: a physics toy, an FPS and a MOBA. Here is what you get, and how to prompt for it.

By GlanceHall · Published October 22, 2026 · 2 min read

Is it possible to vibe code a game? We tested it the strict way: one prompt per game, one attempt, no follow-up messages and no edits. Four AI models each vibe coded three games, a physics toy, a 3D shooter and a one-lane MOBA, and every one of the twelve is playable below.

One-lane MOBA · same prompt, one shot, unedited
Blind testOpen side by side
Being built right now
Claude Opus 5.5
Claude Fable 5.1
Open Claude Fable 5.1’s build
GPT-6 Astra
Open GPT-6 Astra’s build
GPT-6.1 Sol
Open GPT-6.1 Sol’s build

Four vibe-coded MOBAs from one prompt. Open one and click to move; Q W E R cast abilities.

What does a vibe-coded game look like?

A complete prototype with a recognisable style. Because the prompts allowed no image or sound files, every game generates its world in code: low-poly geometry in three.js, textures drawn on a canvas, sound synthesised in the browser. The results range from Claude Opus 5.5's neon city shooter to GPT-6 Astra's forest MOBA lane, and they all start, play and end.

The two GPT models shared more than a style. Their shooters are named Dead Signal and Deadlink and both put a door marked Relay in front of you, and their MOBA heroes both end on an ultimate ability called Starfall, as, it turns out, does Claude Fable 5.1's. Some choices seem to belong to every model trained on the same games, so if you vibe code, expect familiar names unless you ask for new ones.

How long does it take to vibe code a game?

Between 8 minutes and two and a half hours, depending on the game and the model. The physics toy took the GPT models about 8 minutes each. The shooter took GPT-6.1 Sol 13 minutes and GPT-6 Astra 20, while Claude Fable 5.1 took 75 and Claude Opus 5.5 was still refining its game when our 150-minute limit arrived. At list API prices the GPT games cost about $0.65 to $7.27 each.

Balls in a spinning hexagon · same prompt, one shot, unedited
Blind testOpen side by side
Claude Opus 5.5
Open Claude Opus 5.5’s build
Claude Fable 5.1
Open Claude Fable 5.1’s build
GPT-6 Astra
Open GPT-6 Astra’s build
GPT-6.1 Sol
Open GPT-6.1 Sol’s build

The quickest vibe-coded game: twenty balls in a spinning hexagon, physics from scratch.

Black and white photo of a modern workspace with various tech gadgets and peripherals.
Photo by Paul D. Zinn, Jr. on Pexels.

How do you prompt for a game that works?

Be precise about the game and loose about the art. Our shooter and MOBA prompts specify the genre, the controls, the systems such as waves, minions and towers, the camera, and the win and lose conditions, plus the technical rules: three.js provided as a local file, no network requests, every asset made in code. They say nothing about colours or themes, and that is where each model's personality shows.

Then play before you prompt again. A vibe-coded game's first version tells you what the model understood; the second prompt should fix what it missed, not restate the whole game.

Four vibe-coded shooters, names hidden. Which one would you keep playing?

Take the blind test
A man wearing headphones works on a computer with two monitors in a modern office setup.
Photo by Jonathan Borba on Pexels.

Frequently asked questions

Is it possible to vibe code a game?

Yes. Every one of the twelve games here was produced from a single prompt with no human edits, and every one is playable. What you get is a complete prototype; turning it into a game people play for hours still takes iteration.

What AI makes games best?

For the most atmospheric result from one prompt, the Claude models; for fast prototypes you can iterate on cheaply, GPT-6.1 Sol. All four models we tested produced playable games in all three genres.

How should I prompt an AI to make a game?

Describe the genre, the controls, the core systems, the camera and the win and lose conditions, and state the technical rules, such as which engine file to use and that all art and sound must be generated in code. Our game briefs are public templates.

How do you spot a vibe-coded game?

Often by its defaults: procedural low-poly geometry, synthesised sound, a tidy HUD and an arena rather than a level. The games here share those traits because the prompts asked for code-generated assets.

Judge the builds yourself.

Eight briefs, four models, nothing edited. Watch them side by side or take the blind test.

Watch the buildsHow we test

Keep reading

Can AI build it?Vibe-coded websites: what four models made from one sentenceVibe-coded websites from one loose prompt: invent something and make its site memorable. What Claude and GPT models built, how the prompt shaped them, and how to vibe code a site yourself.October 22, 2026Can AI build it?Vibe-coded apps: a phone finance app and a kanban board, four modelsVibe-coded apps you can actually use: a phone finance app with a working send-money flow and a team kanban board with drag and drop, each built by four AI models from one prompt, unedited.October 22, 2026Can AI build it?Claude as a website builder: the site it built from one prompt (open it)Can Claude build a website? Claude Opus 5.5 and Claude Fable 5.1 each built the same marketing site from one prompt. Open both, see what worked, and learn which Claude to use.October 22, 2026