Originally published on VibeGamer, where you can play all three builds in your browser.
We gave Claude Fable 5.2 and GPT-6.2 Astra the same prompt for a small 3D browser game, let each run once with no follow-up messages, and put what came out on VibeGamer unchanged. Gemini 4 Neon gets the same prompt as soon as our Google AI Studio access is sorted out.
Short version: both models made a game that works on first open, with the whole loop the prompt asked for: start, steer, jump, collect orbs, crash, game over, best score and restart. GPT-6.2 Astra in Codex finished in under 4 minutes with a 20 KB file that looks clean, but it faked the 3D in a 2D canvas. Claude Fable 5.2 in Claude Code took 9 to 13 minutes and built real 3D with Three.js and more effects. Its first run, which had no shell and could not test itself, buried the bike in glow. The second run, with a shell like Codex had, tested itself in headless Chrome and the bike is clearly visible. One attempt per model, so treat this as a snapshot, not a ranking.
The three builds at a glance
| Claude Fable 5.2, run 1 | Claude Fable 5.2, run 2 | GPT-6.2 Astra | Gemini 4 Neon | |
|---|---|---|---|---|
| Tool | Claude Code | Claude Code | Codex | Google AI Studio |
| Could run shell commands | No (read and write files only) | Yes | Yes | |
| Time | 9.4 min (567 s) | 12.6 min (758 s) | 3.8 min (226 s) | |
| Agent turns | 22 | 15 | 1 (Codex counts the whole run as one turn) | |
| Follow-up prompts from us | 0 | 0 | 0 | |
| Output | index.html, 45 KB | index.html, 42 KB | index.html, 20 KB | |
| Rendering | Three.js 0.160: real 3D, bloom, fog, particles | Three.js 0.160: real 3D | Canvas 2D with a pseudo-3D projection (no Three.js) | Testing — Google AI Studio access pending |
| Playable on first open | Yes | Yes | Yes | |
| Full loop (start, steer, jump, orbs, crash, game over, best score, restart) | Yes | Yes | Yes | |
| Tested its own game | No: every shell command was blocked, and it said so | Yes: headless Chrome, phone size, a bot run | Partly: a scripted logic test in Node; no visual playtest | |
| Main issue we saw | Bloom washes out the bike, so it is hard to see | START did nothing in our headless Chrome, works in real Chrome | The "3D" is faked in 2D | |
| Cost as reported | $3.62 equivalent (Claude Code's figure; paid by subscription) | $3.27 equivalent (Claude Code's figure; paid by subscription) | About 277K tokens: 268K input, mostly cached, and 9.3K output (subscription) |
The two Fable runs are the same model, prompt and tool; only the permissions differ. All times are wall-clock, from start to the model's final message.
The prompt
Every model got exactly this text, once:
Make a complete, playable 3D browser game as one self-contained file named index.html (you may load Three.js from a CDN; no other files). Game: "Neon Canyon Run" — you pilot a glowing hover-bike racing forward through a neon canyon at night. Steer left/right (and jump) to dodge rock pillars and barriers, and collect energy orbs for points. Speed rises over time. Show score, best score (saved in localStorage) and a speed meter. Crash = game over screen with a Restart button. Controls: arrow keys or WASD plus Space to jump on desktop; on phones, touch the left/right half of the screen to steer and swipe up to jump. Make it look good (lighting, fog, glow, particles) and feel good (smooth steering, camera shake on crash, simple sound effects with the Web Audio API). It must work just by opening index.html in a browser. How we ran it
- One attempt per model, no follow-ups. We sent the prompt, waited for the model to say it was done, and stopped. No "please fix the bike", no second tries.
- Each model in its own agent tool. Claude Fable 5.2 in Claude Code (command line), GPT-6.2 Astra in Codex (command line, workspace-write sandbox, shell allowed). Gemini 4 Neon will run in Google AI Studio.
- Permissions were not equal in the first round. Claude Code was limited to reading and writing files, with no shell, so Fable could not open its own game in a browser. Codex could run commands. To make it fair, we ran Fable a second time with a shell (plus file search), the same kind of access Codex had.
- No edits afterwards. The files you play below are byte-for-byte what the models wrote.
- Costs are what the tools reported. We ran both on subscriptions, so nobody was billed per token. Claude Code shows a dollar figure at API list prices; Codex shows token counts.
Claude Fable 5.2, run 1: real 3D, no way to test it
Fable built the game in Three.js 0.160, loaded from a CDN as the prompt allowed: low-poly canyon walls with glowing lines, a scrolling neon grid, fog, bloom, stars, speed streaks, an engine trail, and an explosion with camera shake on a crash. It was the most visually ambitious of the three.
It also tried to check its work. It tried five shell commands, including a syntax check in Node and a headless Chrome run, and all were blocked. Its final message said plainly that the game was untested and might need fixing. It did work on first open, but the bloom is so strong that the bike becomes a white smear, which makes dodging harder than it should be.
Beyond the prompt it added full-width gates you must jump, a synthesized music loop, pause (P or Esc, and automatically when the tab is hidden) and mute (M or a button).
Play neon-canyon-run-claude-fable-5-2 on VibeGamer
Claude Fable 5.2, run 2: same prompt, with a shell
With a shell, Fable spent longer (12.6 minutes) and used it to test. It ran the game in headless Chrome, checked rendering on desktop and at phone-portrait size, played a no-input run through to game over and the saved best score, and ran a simple dodging bot that survived 150 seconds at top speed on five runs, which shows every row of obstacles can be passed. It said what it could not test: sound, real touch input and how it feels.
The result keeps the real 3D, and this time the bike is clearly visible. Extras: low barriers you can jump, full-width barriers you must jump, a bass loop, a mute button whose setting is remembered, and Space or Enter to start. One oddity: START did nothing in our own headless Chrome (software rendering), but it works in a normal Chrome window.
Play neon-canyon-run-claude-fable-5-2-shell on VibeGamer
GPT-6.2 Astra: fast and clean, but 2D
Astra finished in 3.8 minutes with a 20 KB file, less than half the size of Fable's. It did not use Three.js. It drew the canyon, road, bike and sunset on a 2D canvas with its own perspective projection, so the game needs no internet connection to load. It looks tidy and readable: a synthwave sunset, low-poly rocks, a clearly visible bike, and a HUD with score, best, speed and an orb counter. But it is a 2D drawing made to look 3D, not the 3D game the prompt asked for.
It also tried to test. A headless Chrome screenshot failed, and it was not allowed to open Chrome through computer use, so it wrote a Node script that ran the game's logic: steering, orb pickup, scoring, jumping, pause and resume, crashing, saving the best score, restarting, and touch steering and swipe, plus drawing at three screen sizes. All of that passed. It said that no visual or audio playtest was done.
Beyond the prompt it added pause and mute buttons, an on-screen key legend, and 250 points per orb.
Play neon-canyon-run-gpt-6-2-astra on VibeGamer
Gemini 4 Neon: pending
Gemini 4 Neon will get the same prompt in Google AI Studio, once, with no follow-ups. We have not run it yet because our AI Studio access is still pending, so there are no Neon results here and we won't guess at them. For what Neon has built so far and how it compares on price and benchmarks, see Gemini 4 Neon vs Opus 5.5 vs Fable 5.2 and games made with Gemini.
What this test can and can't tell you
- One run each. The same model with the same prompt can produce a different game next time. Run 1 and run 2 of Fable show how much changes between two attempts.
- Different tools and permissions. We tested Claude Code and Codex as much as the models inside them. Shell access changed what Fable could do, and Codex could not open a real browser either.
- Costs are not comparable bills. Claude Code reports a dollar figure at list price; Codex reports tokens. We paid neither per token.
- "Works" means we played it in Chrome on a desktop. We did not test every phone.
- We picked one kind of game. A small arcade runner in one file. A long project with many files might favour a different model.
Verdict so far
- For a quick, clean prototype: GPT-6.2 Astra. It was 2.5 to 3.4 times faster, the file is small, and it was easy to see and play. If you need real 3D, ask for it explicitly, because it chose a 2D canvas.
- For real 3D and more polish: Claude Fable 5.2, as long as it can run commands. Without a shell it could not see its own bloom problem; with one, it tested the game, and the result kept the real 3D with the bike clearly visible.
- For any model: let it run and test the game. In this test, the run that could open its own game in a browser was the only one that checked how it actually looked.
So what is the best AI for making games? From one prompt each, no single answer: Astra for speed, Fable for 3D, and Gemini 4 Neon still to come. We will add Neon's build and timings here when it has run.
More games from these models: Fable 5.2 games you can play online and GPT-6.2 games.
FAQ
Which is better for making games, Claude Fable 5.2 or GPT-6.2 Astra?
In our one-prompt test, both made a playable game with every feature we asked for. GPT-6.2 Astra was much faster (3.8 minutes against 9.4 and 12.6) but drew fake 3D on a 2D canvas. Claude Fable 5.2 built real 3D with Three.js; its first run had a glow problem that its second run, which could test itself, did not have. One attempt each is not enough to call a winner.
What is the best AI model for game development?
It depends on the game and on the tool. For small browser games in one file, both Fable 5.2 and GPT-6.2 Astra can do it in one prompt. Giving the model a way to run and test the game mattered as much as the model in our test. For a wider comparison with Opus 5.5 and Gemini 4 Neon on price and benchmarks, see Gemini 4 Neon vs Opus 5.5 vs Fable 5.2.
How did Gemini 4 Neon do?
We don't know yet. Neon will get the same prompt in Google AI Studio, and its result will be added to this page. Our access is still pending, so we have not run it.
How much did each game cost to make?
Claude Code reported $3.62 for Fable's first run and $3.27 for the second, at API list prices. Codex reported about 277K tokens for GPT-6.2 Astra (268K input, mostly cached, and 9.3K output) and no dollar figure. We used subscriptions for both, so we did not pay these amounts per token.
Did you fix anything in the games?
No. We sent no follow-up prompts and made no edits. The files you can play are exactly what each model wrote.
Can I use the same prompt?
Yes. It is quoted in full above. Paste it into Claude Code, Codex, Google AI Studio or any other tool, and open the index.html it makes in a browser.
Tested on 21 November 2026 with Claude Fable 5.2 in Claude Code and GPT-6.2 Astra in Codex; Gemini 4 Neon pending. Claude and Fable are trademarks of Anthropic, GPT is a trademark of OpenAI, and Gemini is a trademark of Google. VibeGamer is independent and not affiliated with any of them.
More playable AI-made games at VibeGamer.

