AI game development with frontier models: can one prompt make a playable game?

What does AI game development mean in 2026?

Two different things that share a name. One is AI inside games: non-player characters, procedural worlds and adaptive difficulty, a field as old as game development itself. The other, and the subject of this page, is AI that writes games: a model in a coding agent turning a written brief into the code, art and sound of a playable prototype.

The second kind has moved fast. In the GDC 2026 State of the Game Industry report, based on more than 2,300 industry professionals, 36% said they use generative AI tools at work, 30% of those at game studios. The most common uses were research and brainstorming (81%), daily tasks and code assistance (47% each) and prototyping (35%), and the most used tool was ChatGPT (74%).

Can one prompt make a playable game?

Yes, and here are twelve. We gave Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra and GPT-6.1 Sol three game briefs each, one attempt per game, with a 150-minute limit and no follow-up messages. Every game had to run from local files, use the provided three.js engine and generate all of its models, textures and sounds in code.

Every one of the twelve has a working game loop, input handling, opponents, a HUD and an end or restart state, and none logged a runtime error in our recordings. The MOBAs include minion waves, towers, an AI enemy hero and four abilities each. Nothing was fixed by hand.

What does AI handle well in game development?

The systems a brief can describe precisely.

• Game loops and state. Waves, reloading, health, respawns, scores and restart screens worked in every game.
• Procedural assets. With no asset files allowed, every model built its world from code: low-poly forests, neon cities, rooftop arenas, glowing hexagons.
• Synthesised sound. Every shooter and MOBA generates its sound effects with the Web Audio API rather than recorded files.
• Interface. HUDs, minimaps, kill feeds and, in both Claude MOBAs, a working shop: Claude Opus 5.5's sells items with names like Aegis Plate and Kingsbane, and Claude Fable 5.1's tracks gold and creep score as you play.
• Physics from scratch. All four hexagon simulations keep twenty balls inside a wall that turns once every ten seconds, with no physics library.

What does AI still get wrong?

The parts that come from playing, not from specifying.

• Balance and difficulty. Our scripted input pushes each hero up the lane. Claude Fable 5.1's hero is executed by an enemy tower at 56 seconds, and both GPT heroes fall during the push. In the shooter, Claude Opus 5.5's player dies at 25 seconds and GPT-6.1 Sol's at 35. Nobody tuned these numbers, because nobody played the games before they were finished.
• Originality. Three MOBA heroes from two different companies end on an ultimate ability called Starfall, both Claude MOBAs are called Riftline, and two shooters put a door marked Relay in front of you. Models reach for the same genre conventions.
• Scope control. The most ambitious model, Claude Opus 5.5, ran out of our 150 minutes polishing both its shooter and its MOBA. The GPT models never came close to the limit, and never tried as much.

How do you judge an AI-built game fairly?

By playing every game the same way. Games are the hardest builds to compare, because a slow machine or a lucky moment can make one look better than another. We record every game frame by frame on a virtual clock at 60 frames per second: the recorder controls the game's timers and frame callbacks, so a slow recording machine cannot make one game look choppier than another, and random numbers are seeded identically for every build. The same input script then plays into each game after pressing Play, 40 seconds of moving, turning, firing and reloading in the shooter and 60 seconds of clicking up the lane and casting abilities in the MOBA. If a game reaches its end screen, the recorder presses restart. The methodology page documents the whole process.

That still leaves the most important test to you: open a game and play it. A recording shows what happened under a script; only your hands tell you whether it feels good.

How long does it take, and what does it cost?

From minutes to hours, and from cents to tens of dollars. The physics toy took about 8 minutes for the GPT models and 25 to 40 for the Claude models. The MOBA took GPT-6.1 Sol 20 minutes, Claude Fable 5.1 131 and Claude Opus 5.5 the full 150. At list API prices, a GPT game cost $0.65 to $7.27 and a Claude game roughly $9 to $25. The table shows the MOBA brief in full.

Game builds are the most expensive briefs on the site because they are the largest: Claude's shooters and MOBA run to nine files or more of source, and the agents test them by actually running the game in a headless browser, many times.

How should game developers use AI today?

As a prototyping partner, not a studio. Three ways it already pays off, on this evidence:

1. Paper prototypes that play. Describe the mechanic, generate two or three versions in an afternoon, and play them before committing a team. Our MOBA prompt is a usable template.
2. Game jams and pitches. A playable vertical slice of a lane, an arena or a puzzle in under an hour is a better pitch than a slide.
3. Tools and systems. Physics, procedural geometry and sound synthesis, the parts AI handled best here, are also the parts that slow small teams down.

Then take over. The moment a prototype is worth keeping, it needs a designer to tune it, an artist to give it a look of its own, and a programmer to restructure code that was written in one sitting. The largest of our MOBAs, Claude Opus 5.5's, spans seventeen script files written in a single session; that is a lot of code to inherit without its author.

What do game developers think about AI?

Increasingly sceptical, even as use grows. In the same GDC survey, 52% of respondents said generative AI is having a negative impact on the industry, up from 30% a year earlier and 18% the year before, with programmers at 59%; only 7% saw a positive impact. Having played these twelve games, I understand both halves of that number. The prototypes are impressive and genuinely useful. They are also, every one of them, unmistakably unfinished, and finishing is still the job.

Frequently asked questions

Can AI make a playable game from one prompt?

Yes. Four frontier models each built a 3D shooter, a one-lane MOBA and a physics toy from one prompt per game, with no human edits, and all twelve games run in a browser with controls, opponents, a HUD and an end state.

Is AI used in game development today?

Increasingly. In GDC's 2026 State of the Game Industry survey of more than 2,300 professionals, 36% said they use generative AI at work, most often for research and brainstorming, code assistance and prototyping.

What can't AI do in game development yet?

Game feel, balance and long-term design. The games here are complete prototypes, but none was tuned by a player: heroes walk into towers, difficulty is untested, and several models even chose the same ability names.

Which AI is best for making games?

The Claude models built the richest prototypes, the GPT-6 models the fastest and cheapest. Our best-AI-for-game-development page compares all four with times and costs.