The best AI for game development is the one whose prototype you would keep playing. So we asked four frontier models for three games each, a 3D shooter, a one-lane MOBA and a physics toy, one prompt per game and no edits. All twelve games are playable below. The short answer: Claude builds the richer game, GPT builds the faster one.
Which AI is best for game development?
It depends on which constraint bites. If you want the most complete prototype from a single prompt, use a Claude model: their games have more lighting, effects and systems, such as shops, minimaps and creep-score counters. If you want to try ten ideas this afternoon, use GPT-6.1 Sol: its games arrived in 8 to 20 minutes at about a dollar each at list prices.
| Model | Finished | Wall time | Output tokens | Files | Source size | Runtime errors | API-equivalent cost | Attempts |
|---|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 | Yes | 2h 30m | not recorded (stopped at the 150-minute limit) | 17 | 286 KB | 0 | — | 2 |
| Claude Fable 5.1 | Yes | 2h 10m | 276k | 14 | 204 KB | 0 | $24.76 | 4 |
| GPT-6 Astra | Yes | 24m 47s | 74k | 5 | 95 KB | 0 | $7.27 | 1 |
| GPT-6.1 Sol | Yes | 19m 41s | 67k | 3 | 85 KB | 0 | $1.12 | 1 |
What did each AI get right?
Different things.
- Claude Opus 5.5 built the most atmospheric shooter, a neon city at night, and the densest MOBA, with a shop, a minimap and a named enemy hero, and used our full 150-minute limit on both.
- Claude Fable 5.1 finished both of its big games inside the limit: a dusk rooftop shooter whose player survives our whole script, and a forest MOBA with a shop and a river crossing.
- GPT-6 Astra built a lush low-poly forest MOBA lane in 25 minutes and a shooter whose player survives the script.
- GPT-6.1 Sol built every game fastest and cheapest, in a darker, sparser style.
What did none of them get right?
Game feel and balance, the parts that take playtesting. Under our fixed input, several heroes from both vendors die, Fable's MOBA hero walks into a tower and both GPT heroes fall in the push. None of that is a bug; it is a game that has never been played by a person. Every model also reached for similar fantasy names: three of the four MOBA heroes end on an ultimate called Starfall, and both Claude MOBAs are called Riftline. Treat an AI game as a playable design document, then iterate.
How should you use AI for game development?
Prototype with it, then take over. Our MOBA and shooter prompts show the level of detail that works: genre, controls, systems, camera, end conditions and technical rules, with the art left to the model. Generate two or three prototypes, play them, and carry the best ideas forward. For a wider look at where AI fits in a game studio, read AI game development.
Four shooters, names hidden. Play each for a minute and pick the one you would keep building.
Take the blind test
Frequently asked questions
What is the best AI for game development?
For the most complete, atmospheric prototype from one prompt, Claude Opus 5.5 or Claude Fable 5.1. For fast, cheap prototypes you can iterate on, GPT-6.1 Sol. All four models we tested built three playable games, so the choice is about ambition, time and cost.
Can AI make a full game?
It can make a complete, playable prototype: every game here has a working loop, controls, enemies or opponents, a HUD and an end state. A finished commercial game still needs design iteration, balance, content and polish that one prompt does not supply.
Which AI is best for game dev with three.js?
All four models built working three.js games from the same prompt. The Claude models produced richer scenes and effects; the GPT models produced simpler scenes much faster. Every build uses the same three.js release, provided as a local file.
How much does it cost to make a game with AI?
At list API prices, our GPT game builds cost $0.65 to $7.27 each and the Claude game builds roughly $9 to $25. On a subscription, the cost comes out of your plan's usage limits instead.