Inside Korea’s AI Game Boards: The Builds, the Know-How, and What Keeps Breaking
A Warhammer 40K survivors game pulled 95,726 views on FMKorea; an Uma Musume autochess pulled 73,580 and was gone thirteen hours later. Opened and checked on 13 September - plus the reasoning-tier economics, the four-tool music pipeline, and the costs nobody mentions until the server dies.
TL;DR — Two Korean forums have turned into an AI game-making workshop. A Warhammer 40K survivors game drew 95,726 views; an Uma Musume autochess drew 73,580. I opened all of them on 13 September: the Warhammer one plays, and the Uma one was already gone, thirteen hours after its post. Underneath the games is real, transferable know-how — which reasoning tier is worth paying for, how people are getting original music out of a text model, and the operating costs nobody mentions until their server dies.
All three opened on 13 September 2026. Left to right: The Last Litany, Sanctuary Survivors, Astra Attack.
The English-language coverage of GPT-6 Astra games points at two GitHub repositories. That is not where the volume is. Two Korean boards — FMKorea's Digital/AI section and DCInside's AI-utilisation gallery — are producing playable games faster than anyone is cataloguing them, and the comment threads underneath carry more practical detail than the repos do.
This continues the earlier round-up; everything here was opened and checked on 13 September.
The games
The Last Litany — Warhammer 40K survivors
last-litany.edusub-mail-7720.chatgpt.site · FMKorea 9 Sep · 95,726 views, 170 recommends, 136 comments · live
A Vampire-Survivors-style game set in the Black Templars chapter. Nine weapons, six part-resonances, three fronts. The creator describes himself as a Warhammer player who collects miniatures, says he built it after seeing another survivors game reach the front page, and asks people to try it "while you're in the bathroom" and report back.
That request produced the most useful thing in the thread. A player reported a specific bug — stalling to full-farm a stage broke progression — and the creator replied "수정했습니다" (fixed) roughly seven hours later. The forum thread is the bug tracker.
Sanctuary Survivors — Diablo II survivors
dia2.hardt.in · posted in the Last Litany comments · live
Not a front-page post — it arrived as a reply, from a reader who had made his own. Character select, a five-slot active skill bar with Fire Ball / Nova / Blizzard, a skill tree, equipment forging, gold. The polish level is the thing to notice: this is somebody's comment, not somebody's launch.
Astra Attack — dead three days, then resurrected
scar-flame.vercel.app · live again
The 4v4 browser FPS from the earlier article. When I checked it on 10 September it returned 503 on three consecutive probes. It had run out of server.
On 12 September the creator posted a relaunch, with the resurrection framed in mock-scripture: a moderator "came forward and pleaded, saying I am able to offer up a server, I beg you, open the door again", and so "three days after the breath had stopped, he rose again from the dead."
What changed, in his own list: the Warehouse map is gone (the most popular one — he says so and apologises), several new maps replace it for now, new guns, UI cleanup, working mobile. He has opened donations, and says plainly that they probably will not cover the server bill. He also asks, again: "please stop using hacks."
The line that matters is his own limit statement — he will keep patching "as far as time, money, and the Astra usage cap will stretch."
Uma Fight Tactics — the one that is already gone
FMKorea 13 Sep 00:14 · 73,580 views, 152 recommends, 135 comments · down
An Uma Musume × autochess game, built to resemble Teamfight Tactics. The post shows a main menu, a collection screen, motion viewer, options, season select, and in-game combat. The creator states his own asset split: "BGM from Suno, images from GPT."
Thirteen hours after that post, umafighttactics.onrender.com serves this at the root:
{"service":"multiplayer","endpoint":"/multiplayer"}
Every path I tried — /index.html, /game, /play, /app, /multiplayer — returns {"service":"multiplayer","error":"Not found"}. The signalling service is up; the game is not deployed behind it. Whether that is a misconfigured redeploy or a takedown, I cannot tell from outside.
Two of the four biggest community games this week have been unreachable at some point in the last four days. That is the actual state of this scene, and no list that does not check will tell you.
The know-how
This is the part worth taking away, and it is buried in threads rather than written up anywhere.
1. Which reasoning tier to pay for
A DCInside post titled "Astra efficiency comparison" (2,488 views) makes a claim worth checking: that Medium and High are the efficiency sweet spot and the rest sit in shadow. Checking it against Artificial Analysis on 13 September:
| Tier | Intelligence | Output speed | Cost per task |
|---|---|---|---|
| max | 53 | 60 t/s | $3.26 |
| xhigh | 53 | 51 t/s | $2.31 |
| high | 51 | 52 t/s | $1.72 |
| medium | 50 | 50 t/s | $1.54 |
| low | 46 | 54 t/s | $0.82 |
Two things fall out that the marketing does not say.
max and xhigh score identically — 53 — and max costs 41% more. If you are paying for max on the assumption it thinks harder, the benchmark does not support that; you are buying output speed.
high (51) costs about half of max (53). Two points of intelligence for double the money. For generating a browser game, where the bottleneck is your own iteration count rather than the model's ceiling, medium or high is the rational default — which is exactly what the gallery poster concluded, and he was right.
Prices vary up to 4× across the range. On a hobby project where you will run dozens of passes, that is the difference between finishing and hitting your cap — the constraint the Astra Attack creator named explicitly.
2. Getting original music out of a text model
The most inventive workflow this week, posted as a "fine-dust tip" (1,178 views). The pipeline:
- Install REAPER and connect it over MCP
- Have Astra fetch an existing game's BGM, analyse it, and compose something in a similar vein through the MCP connection
- Accept that most of it is bad
- From roughly 90 seconds of output, about 10 seconds will be genuinely usable — take only that fragment
- Feed the fragment into Suno as a mashup / sample
- Ask Gemini to write the Suno prompt
The yield figure is what makes this credible. A tip that claims a text model composes good music would be worthless; one that says one part in nine is worth keeping, here is how to harvest it is a working process. It is also four tools deep, which is a fair picture of what "made with AI" means in practice.
3. Anchor on a game everyone knows
Every successful game here is a named clone: Vampire Survivors, Teamfight Tactics, Diablo II, Sudden Attack. The Uma creator's stated instruction was to make it "as close to TFT as possible."
This matches what the published one-shot prompts do — point at a game whose rules are common knowledge, so the specification comes free and the model spends its effort on execution instead of design. Nobody in these threads is asking for an original genre.
4. Split the assets across tools
Nobody is generating everything in one place. The stated splits: images from GPT, music from Suno, code from Astra, and in the BGM case a fourth model writing the prompt for the third. The text model is the engine; it is not the studio.
What actually breaks
The failures are consistent, and none of them are about code generation.
Server cost. Astra Attack died for three days and came back only because a reader donated hosting. Its creator has opened donations he expects to fall short. Uma Fight Tactics is down with its multiplayer service still running — a deployment problem, not a programming one.
Cheating. "Please stop using hacks" is now in the Astra Attack launch post itself. A browser game with no anti-cheat gets one within a day.
The usage cap. "As far as time, money, and the Astra usage cap will stretch" is the real development schedule. At $3.26 a task on max against $1.54 on medium, tier choice is the schedule.
Nothing about rendering or code. Not one thread complains that the model could not build the thing.
The part nobody has settled
Every game above is built on somebody else's intellectual property. Warhammer 40,000 and the Black Templars belong to Games Workshop. Diablo II and Sanctuary belong to Blizzard. Uma Musume belongs to Cygames; Teamfight Tactics to Riot.
The Uma thread's top comment puts the objection directly: this is an existing autochess system with famous characters skinned onto it by AI, and "if they put a single ad on it I'll report it — I've saved everything."
It sits at 27 up and 25 down. The community has not resolved this, and the split is close to even.
The line the objection actually draws is monetisation, not creation — which is roughly where fan projects have always sat. What is new is throughput. A fan game used to take a team a year, which functioned as a filter. At a few hours each, the filter is gone, and rights holders have not yet responded to the volume. Note that one of the two games already unreachable is the one with the trademarked characters and the loudest complaint thread; I have no evidence those facts are connected, and I am not asserting they are.
What it changes
The most quoted line in the Uma thread is not about the game:
When the execution tools flatten out, the idea itself becomes the competitive edge.
The week supports it. The Diablo II game arrived as a comment. Somebody read a post about a survivors game and shipped their own before the thread cooled. The distance between seeing a thing and having built one is now hours.
What has not flattened is everything after the build: hosting, cheating, deployment, rights, and the will to keep patching when nobody is paying you. That is where all four of these projects have spent their difficulty.
What this article does not establish
Reach figures are single readings taken on 13 September and will have moved.
I opened each game and confirmed what loads; I did not play through them, and I cannot verify any creator's account of how they were built. "Made with Astra" is the author's claim in every case.
The Uma game's disappearance is observed, not explained. I checked six paths and reported what I got. It could be a redeploy in progress.
The efficiency table is Artificial Analysis's benchmark, not mine, and benchmark intelligence scores are a poor proxy for whether a model can build your specific game. The useful part is the price ratio, which is not a matter of opinion.
FAQ
Where are Koreans posting AI-made games?
FMKorea's Digital (컴퓨터/폰/IT) — AI board, and DCInside's AI-utilisation minor gallery. Both are producing playable browser games weekly, with the creators answering bug reports in the comments.
Which GPT-6 Astra tier should I use to build a game?
Medium or High. Per Artificial Analysis on 13 September, max and xhigh score the same intelligence (53) while max costs 41% more, and high scores 51 for about half the price of max. On a project where you will run many passes, the cheaper tier buys more iterations, and iterations are the actual bottleneck.
How are people making original game music with AI?
The workflow shared this week: connect REAPER over MCP, have Astra analyse an existing track and compose through the MCP, keep only the ~10 usable seconds out of ~90, feed that fragment into Suno as a sample, and have Gemini write the Suno prompt. Four tools, and the honest yield is roughly one part in nine.
Are these games safe to play?
They are browser games on personal hosting with no anti-cheat and, in at least one case, known cheating. Treat them as hobby projects: no accounts, no payment details, and expect them to disappear. Two of the four here have been unreachable in the past four days.
Can I get in trouble for making a fan game with AI?
This article cannot give legal advice, and the Korean threads have not settled it either — the top objection sits at 27 up against 25 down. The distinction people are drawing is monetisation: a free fan project attracts different attention from one carrying ads. Every game described here uses trademarks belonging to Games Workshop, Blizzard, Cygames or Riot.
← Back to all posts