What people actually build when AI writes the game
When the cost of trying an idea collapses, what do people try? This is what sixty nine finished, published games turned out to be.
Jump to a section
Arcade and action dominate. 3D is the third most common tag, which almost nobody would have guessed. Only about a third work on a phone, and fewer than one in ten have any online play. The games that grew beyond a single file cost roughly twenty three times more to build, without being played more.
Sixty nine published games on Agent Games, read from the catalogue on 27 October 2026. Creators enter their own figures when they publish, so treat every number as an honest approximation. The set is also self selected: these are games somebody finished, and nothing here counts an abandoned one.
What gets built#
The arcade and action lead is unsurprising. A short run with a score is the shape that suits a browser, and it is the shape every guide recommends for a first build.
The surprise is 3D sitting third, ahead of puzzle, platformer and racing. Three dimensions used to be the thing you worked up to over years. It is now something people reach for in week one, apparently because the barrier was never conceptual, it was the amount of code you had to write before anything appeared on screen.
Survival placing that high is the other oddity. It is a genre of overlapping systems, which is exactly what used to make it a bad first project.
The finding worth arguing about#
Roughly three in ten finished games here are still a single HTML file. The other seven grew into small projects, averaging about eighteen files and a megabyte and a half.
Those projects averaged 8.8 million tokens against 385 thousand for the single file games. That is a twenty three fold difference in what it took to build them, and it bought no measurable increase in how often they are played.
Plays depend heavily on when a game was published, how it was promoted and whether it reached the front page. This comparison does not control for any of that, so it is not evidence that ambition is wasted. What it does show is that ambition is not automatically rewarded, which is a different and more useful claim.
Two thirds of them do not work on a phone#
Twenty five of sixty nine are marked as playable on a phone. That is the number that should bother anybody publishing a browser game, because most people who follow a link to one are holding a phone when they do it.
It is also the cheapest gap in this entire article to close. A keyboard only game usually needs one CSS rule and one touch handler to become playable with a thumb, which is a single prompt rather than a rewrite.
| Feature | How many of the sixty nine | What it suggests |
|---|---|---|
| Works on a phone | 25 | The largest easy win available to most creators. |
| Has a leaderboard | 54 | Score chasing is the default shape, and it suits the medium. |
| Plays with other people | 6 | Online play is still rare, and it is the hardest thing here. |
| Has cover art | 68 | Essentially everyone. The one that does not stands out badly. |
Counts of published games on 27 October 2026, not percentages of anything wider.
The leaderboard figure is the one that reveals the shape of the medium. Four in five published games end in a number, because a score is what makes a browser game worth opening twice. Games that do not end in a number are harder to make and harder to leave.
Cover art tells a different story. Almost every game has it, because asking for one is a single request and takes a minute. That is the pattern across this whole data set. Anything a model can do unprompted is done nearly everywhere, and anything that needs a person to notice something is done by a minority.
What this says about building with AI#
Three things, offered as readings rather than as conclusions.
- The hard genres stopped being hard to start. 3D and survival both place higher than their difficulty would once have allowed. Starting is no longer the filter it was, which means finishing is now almost the whole filter.
- Scope creep has a measurable price. The difference between one file and eighteen is not a little extra work. In this set it is most of the entire build cost.
- The last ten percent is still the last ninety. Nearly everyone produced cover art and a leaderboard, because those are one request each. Far fewer made the game work on a phone, because that requires playing it on one and noticing.
📊The catalogue these numbers came from
Every page lists the model, the prompt and what the build took.
Fable 5.1
GPT-6.1 SolCommon questions#
What kinds of games do people make with AI?
Arcade and action games lead by a wide margin, followed by 3D and survival. Short runs with a score suit the browser and suit the way these games get built, so that is what most of them are.
Can AI really make a 3D game?
On the evidence here, often enough that 3D is the third most common tag across published games. The barrier was never the concept, it was the volume of setup code needed before anything appeared on screen, and that is the part that has collapsed.
Should I keep my game in one file?
For a first game, yes. Single file games in this set cost a fraction of what multi-file projects cost and are played about as often. Split the file when the game genuinely needs it, not because a project feels more serious.
Do browser games need to work on phones?
Most people following a link to a browser game are on one, and only about a third of the games here work that way. It is the largest and cheapest gap in this data: usually one CSS rule and one touch handler.
How many AI-built games have multiplayer?
Six of sixty nine. Online play remains the hardest thing to build in this medium. It needs a rule about which machine decides what happened, which is a design problem rather than a coding one.
Are these numbers representative of AI game development generally?
No. They describe one catalogue of finished, published browser games, reported by their creators. They are useful as a picture of what people actually choose to build and finish, not as a survey of the industry.
Where to go next#
For what these builds cost in tokens, hours and dollars, with the full distribution, read the cost breakdown. For where AI genuinely fits in making games, AI game development sorts the work by how well it does each part.
If you want to add to this set rather than read about it, the ideas list has twelve sized to one evening, and publishing is free. Fill in the tokens and the minutes when you do, because that is where every number on this page came from.
Built something? Put it on Agent Games.
Publishing takes a file and a minute. Your game gets its own page listing the model, the prompt and the token count, and anyone can play it without signing up.
Read next