Astra Attack: The AI-Built Multiplayer Browser FPS Korea Is Calling "GPT-6 Sudden Attack"

TL;DR — Astra Attack is a real 4v4 / 5v5 multiplayer FPS that runs in a browser tab, built with GPT-6 Astra, playable right now for free with no install and no login. Korean forums nicknamed it "the GPT-6 Sudden Attack" and one repost pulled 252,000 views in a day — enough traffic to knock its server over. It is also the strongest counterexample to the "one prompt, one game" story, and the proof is sitting in its own credits file: documented map ports from two open-source games, reproducible import scripts, pinned upstream revisions and package checksums. Someone did real work here.

Yesterday I wrote about the GPT-6 Astra one-shot game prompts and measured what the outputs actually contain. The most impressive Astra game currently circulating did not come from that scene at all. It came from a Korean forum, it is multiplayer, and it is the least "one-shot" thing on the board.

Play it first

scar-flame.vercel.app — opens and plays in a desktop or mobile browser. Free, no login, no API key. There is a solo shooting range if the multiplayer queue is full, which it often is.

Astra Attack main menu. Screenshot captured 9 September 2026 at 1280×720.

What is actually in it, read off the running build:

Modes	Multiplayer Warehouse 4:4 and ATCS HD 5:5; solo 20 m shooting range
Weapons	SCAR-H rifle (20 rds, semi/auto), M700 bolt-action (5 rds, 3× scope), Mk2 pistol, one grenade
Optics	Iron sights, Aimpoint red dot 1×, EOTech EXPS3 1×, rifle scope 2× — swapped with a hold-and-drag radial
Progression	XP, levels, a public ranking board, persistent profile with matches / wins / kills / deaths
Netcode-facing	Room list, quick join, 3:30 match clock, scoreboard, kill feed, ping readout, entry queue
Mobile	PWA install, landscape prompt, on-screen controls
Scoring	Kill +20, finish +50, win +30 XP; finish and win rewards need 30 s of activity and staying to the end; kills capped at 50 per match

The solo 20 m range. Screenshot captured 9 September 2026. Game by 봄바다; see the in-game credits panel for its full third-party asset attributions.

That is not a tech demo. It is a small game with a progression loop, a matchmaking lobby, and a server bill.

Where it came from

The creator, posting as 봄바다, put it on DCInside's AI-utilisation board on 8 September 2026 at 00:24 under the title "[Astra Attack] 3:3 FPS official open v2." That post has 13,533 views, 49 recommends and 63 comments.

Later the same day, at 16:58, a repost on FMKorea titled "GPT6 서든어택 떴다" ("the GPT-6 Sudden Attack is here") went to the front page. As of 9 September it stands at 252,584 views, 382 recommends and 231 comments.

The nickname needs one line of context for anyone outside Korea. Sudden Attack is Nexon's long-running Korean tactical FPS, and its signature look is a shipping-container warehouse. Astra Attack ships a warehouse map. The comparison is about the setting and the feel, not about copied assets — the asset trail, below, is documented and points somewhere else entirely.

The top comment on the FMKorea thread, with 166 upvotes, is a joke that someone had already built an ESP wallhack for it. A player in the DCInside thread reported being one-shot killed by a suspected cheater within the first day. Faithful to the original, as several commenters put it.

What the creator actually claimed

The whole post body, translated, is a changelog and three lines:

• Paid Tokyo server to lower ping
• Headshot / kill feedback
• Adjustable mouse sensitivity
• Blood effect on hit
• Model fixes
• (two joke items about disinfected chopsticks and wet towels)
• "Token usage: only 30% of Pro-20!!"
• "Mobile support patch tomorrow"
• "The singularity is coming"

Three things in there are load-bearing.

It is v2. Not a first attempt that happened to work. There was a previous release and this is the iteration on it.

"Only 30% of Pro-20." The creator's own boast is about how much of a month's ChatGPT Pro token allowance the project consumed. You do not measure a one-shot in percentages of a monthly quota. That figure is a description of many long sessions.

A paid Tokyo game server. Later in the thread, replying to lag complaints: "too many concurrent players, the server keeps falling over, working on it." And to a request to raise the player cap: "give me some money, servers are expensive." This is an operational project, not an artefact.

One more thing, by absence. A commenter wrote simply "share the prompt." It went unanswered. Another asked what language and engine it was built in. Also unanswered. Whatever produced this, the recipe is not public — which is exactly the opposite of the one-shot repositories, where the prompt is the whole point.

The credits file is the real story

This is the part that convinced me, and it is the part nobody in either forum thread mentioned. The game ships per-map credit files. They read like a careful open-source port, because that is what they are.

Asset	Origin	Licence
ATCS HD map	Advanced Tremulous Combat Simulator — Nicolas "Jex" Jansens, Superpie, Stijn "Ingar" Buys; source port by Thomas "illwieckz" Debesse / Interstellar Oasis	CC BY-SA 2.5 (map), CC BY-SA 2.0 (TRaK textures), CC BY 4.0 (port changes)
Platform 23 map	Jack "EmperorJack" Purvis and the Unvanquished team	CC BY-SA 3.0; PK02 textures CC BY 3.0; sky CC BY-SA 3.0; shader data CC0
Mobile control icons	Lorc and Delapouite, game-icons.net	CC BY 3.0
UI icons	Tabler Icons, Paweł Kuna	MIT
Music ("Dark")	pheonton, via OpenGameArt	CC BY-SA 3.0
M700 model	Stein Games	CC0
Weapon fire audio	areniporgen, via Freesound	CC0
Surfaces	Poly Haven	CC0
Hands / arms	EXUnum	CC BY 4.0
Soldier animation	Adobe Mixamo Vanguard	Mixamo terms

Read the map notes and the level of care becomes obvious. On ATCS HD:

Changes for Astra Attack: all map coordinates initially scaled to 80%, then horizontal area reduced by another 10% (X/Z multiplied by sqrt(0.9)); character dimensions unchanged; BSP conversion, curve tessellation and matching convex collision; four diffuse lightmaps packed into an atlas; WebP textures capped at 1024 pixels; team spawns and Korean location names.

And, further down: a source.zip with the editable map, a scripts/import-atcshd.py you can re-run, a map.json recording the pinned source revision and the package SHA-256, and a note that a separately-licensed texture was swapped out for a compatible one.

That last detail is the tell. Noticing that one image in a large asset pack carries a different licence, and replacing it, is not something that falls out of a chat session. Neither is shipping the editable source to satisfy a share-alike obligation, nor pinning an upstream revision so the import is reproducible.

Whoever built this understood BSP geometry, lightmap packing, collision meshes and copyleft compliance. GPT-6 Astra may well have written most of the code. It did not decide to do any of that.

What broke — and why that is the interesting half

The failures are as informative as the successes, because they cluster in exactly one place.

Capacity. "40 concurrent players, seriously?" one commenter wrote, delighted, at 00:53. Two minutes later the creator was posting that the server was falling over. By 20:24 the next evening, after the FMKorea front page hit: "the server keeps dying because of the FMKorea crowd." The game has a queue system now because it had to grow one.

Netcode. The sharpest comment in the thread diagnoses it precisely: movement stutters and reloads stall because the client waits on the server for every input, instead of predicting locally and letting the server reconcile afterwards. Client-side prediction with server reconciliation is the standard architecture for exactly this reason, and its absence is the single most-reported complaint.

Cheating. No anti-cheat. An ESP was reportedly built within a day, and players reported one-shot kills.

Small UI mistakes with big consequences. Crouch was bound to Ctrl, so Ctrl+W — crouch plus forward — closes the browser tab. Right-click gesture conflicts do the same. Several players lost matches to it before anyone worked out why.

Line those up and they share a property: none of them are visible in a single-player test session, and none of them can be judged by looking at the output. They only appear under concurrent load, adversarial users, and hands that hold keys down. That is a reasonable rule of thumb for where AI-assisted game development currently sits — rendering a convincing 3D world is now cheap; making a shared one behave is not.

So is it a "one-shot game"?

No, and its own author never said it was. That framing was added by the reposts.

Set it against the two-prompt case studies in yesterday's piece:

	Mosswing (one-shot)	Astra Attack
Prompt published	Yes, verbatim	No — a request for it went unanswered
Turns	One	Unstated; "30% of a Pro-20 month" of tokens
Version	First attempt	v2, actively patched
Delivery	One 701 KB HTML file	Vite build, PWA, paid game server
Third-party assets	None (Three.js inlined)	Ten-plus sources with per-file licence notices
Multiplayer	No	Yes — and it is where all the problems are

The most impressive GPT-6 Astra game in circulation is the one that took the most human work. That is not a knock on the model, and it is not a knock on the creator. It is the actual shape of the thing, and it is more useful than either the hype or the backlash.

What I could not verify

I played the solo range and read the shipped build, its credits, its asset files and both forum threads. I did not verify the build process, because no repository, prompt log or session transcript has been published. The "30% of Pro-20" figure is the creator's own claim with nothing attached to check it against. I did not join a ranked match, so the netcode complaints here are other players' reports plus the game's own architecture-facing UI, not my measurement. And the model attribution rests entirely on the creator saying so — the same caveat the community game list applies to every entry it carries.

The game itself is checkable by anyone: open the link.

FAQ

Where do I play the GPT-6 Astra Sudden Attack game?

scar-flame.vercel.app. It runs in a normal browser on desktop or mobile, free, no login. Its real name is Astra Attack; "GPT6 서든어택" is the nickname a Korean forum gave it.

Is it actually multiplayer?

Yes — 4v4 on Warehouse and 5v5 on ATCS HD, on a paid Tokyo server, with a room list, ranking board and persistent XP. If it is full you will get a queue position; there is also a solo shooting range that always works.

Was it made with a single prompt?

No. The creator released it as v2, described the cost as roughly 30% of a month's ChatGPT Pro token allowance, and did not answer a direct request to share the prompt. The single-prompt framing came from reposts, not from the author.

Did it copy Sudden Attack's assets?

Nothing suggests that. Its two multiplayer maps are documented ports of ATCS HD and Platform 23 from the open-source Tremulous and Unvanquished projects, under CC BY-SA, with the editable source shipped and upstream revisions pinned. The resemblance to Sudden Attack is a shared setting — a container warehouse — not shared files.

Why does it stutter?

By the most detailed player diagnosis in the original thread, the client waits on a server round trip for movement instead of predicting locally and reconciling afterwards. Server overload made it worse: the game hit the front page of a major forum and the creator spent the evening restarting it.

What does this say about making games with AI?

That the hard part moved. A single-player 3D world that looks and feels right is now within reach of one person and a model. A shared world — concurrent load, cheaters, latency, input edge cases — still needs someone who knows what they are doing. Every problem this game has is on the second list.

---

← Back to all posts