What People Actually Built With GPT-6 Astra: The Viral Projects, Opened and Checked

TL;DR — Six days after GPT-6 Astra shipped, the most-shared thing built with it in Korea was a browser game that pulled 344,000 views on a single forum post — with no playable link. The one people actually played had, by the time I opened it, a patch archive listing ten versioned releases in under 48 hours, three of them within thirty minutes of each other. The gap between what gets attention and what gets built is the story. Every project below was opened and screenshotted on 9 September 2026; where something could not be verified, it says so.

The community reaction to GPT-6 Astra has been dominated by one Korean word — 딸깍, roughly "one click," used the way English uses "and boom, done." It is a good joke and a bad description. Below are the projects that actually caused a stir, in Korean forums and internationally, with what each one really is.

All four opened on 9 September 2026. Clockwise from top left: Astra Attack, LAST SIGNAL, Arcane Ascent, 湖畔.

1. LAST SIGNAL — the one with a patch archive

last-signal.naitoria.chatgpt.site · posted to FMKorea 8 Sep 08:27 · 236,642 views, 535 comments

A wave-survival shooter in the Vampire Survivors mould, with a global leaderboard. The post title translates as "I'm the guy who one-clicked a Vampire-Survivors game yesterday," and the body is one line: "Updated it. Bug fixes, difficulty tuning — I just told it to and it handled it fine."

Then I opened the game and a patch-notes modal came up.

The in-game patch archive, read 9 September 2026.

Release	Version	Timestamp
UPDATE 01 · RECONNECTION	v1.1.0	8 Sep
UPDATE 02 · SIGNAL RUSH	v1.3.0	8 Sep
HOTFIX 01 · MEDKIT TUNING	v1.3.1	8 Sep
UPDATE 03 · BUILD ARCHIVE	v1.4.0	8 Sep
HOTFIX 02 · SIGNAL CLARITY	v1.4.1	8 Sep
UPDATE 04 · BREAKPOINT	v1.5.0	9 Sep
UPDATE 05 · BEYOND OMEGA	v1.6.0	9 Sep
UPDATE 06 · AFTER THE LAST LIGHT	v1.6.1	9 Sep 12:25 KST
UPDATE 07 · CONTINUOUS SIGNAL	v1.6.2	9 Sep 12:40 KST
UPDATE 08 · SEEKER VECTOR	v1.6.3	9 Sep 12:55 KST

Ten named releases in under 48 hours. Three of them thirty minutes apart. v1.6.3 shipped while I was reading v1.6.2 — I had to re-read the archive to catch it.

And the patch notes are not padding. UPDATE 07 is two items: keep the background music playing while the player picks a level-up card, and tidy up which states stop the audio. That is somebody noticing a small break in feel and fixing it.

The thread explains where the list came from. A player wrote that the ranking system felt wrong — you score more by circling and avoiding the final boss than by killing it. The creator replied asking what to prioritise. The player laid out a design: past a certain difficulty the boss should not end the run, later bosses should keep spawning at shortening intervals, until survival becomes impossible. The creator's answer was that the difficulty scaling was going up substantially.

That is a live-service loop with player-driven balance design, running on a forum thread. Whatever "딸깍" means, it does not mean this.

2. Arcane Ascent — same author, previous day

arcane-ascent-game.naitoria.chatgpt.site · posted 7 Sep 21:36 · 286,159 views, 570 comments

A deckbuilding roguelike — three heroes with distinct passives, ten-card starting decks, three regions per run, a card compendium. The author's own framing was blunter than the game deserves: "asked it for a card game this time. Indie studios are finished."

That line is what set the thread off, and the two highest-voted replies disagree with each other in a way worth preserving. The top comment (373 votes) argues that slop games from people with no eye for design have no effect on studios at all. The second (214) draws a sharper line: what is threatened is the outfit shipping cheap library-filler titles, not indie developers, because indie is a contest of ideas and those do not get automated away.

Both of them are arguing with the author's headline, not with the game — which, played, is a competent small roguelike with AI-generated character art.

3. Astra Attack — the multiplayer FPS

scar-flame.vercel.app · two FMKorea front pages on 8 Sep: 58,477 views at 01:18 and 252,584 at 16:58

A 4v4 / 5v5 browser FPS with a paid game server, nicknamed "the GPT-6 Sudden Attack." I looked at this one closely enough that it has its own writeup — the short version is that its own credits file documents two map ports from open-source projects, with reproducible import scripts, pinned upstream revisions and package checksums, which is not something a chat session decides to do.

It is also the clearest case of the reach pattern: the same URL hit the front page twice in sixteen hours, and the second post did four times the traffic of the first.

4. The League of Legends clone — biggest reach, least to check

FMKorea, 9 Sep 00:02 · 344,076 views, 705 recommends, 368 comments

The largest numbers of anything in this article, by a distance. It is a six-second clip of a MOBA — lane, minions, abilities, an annotation saying multiplayer works — posted with the line "this is a one-click result? insane."

There is no playable link and no repository. So this is the honest entry in the list: the single most-shared GPT-6 build in Korea this week is the one nobody outside the author can verify.

The comment section is a small masterpiece of the form. The top reply, at 1,058 votes, is not about AI at all — it is a player objecting that the bot took the wrong ability at level one. The next two are about the bill: a joke that a bare "HELLO" burns five hours of token quota, and the author's reply that with no context the model just spends tokens guessing what you meant.

5. Smash Karts Arena — the best receipts, the least attention

github.com/amsminn/gpt-6-astra-smash-karts · created 5 Sep · 1 star

An 8-player 3D kart battler: 11 weapons, three-minute free-for-alls, bots, live leaderboard, server-authoritative movement and hit validation, Discord Activity integration. The README claims one prompt and about 42 minutes in Codex CLI, recorded model gpt-6-astra.

What makes it worth singling out is that the claim comes with a receipt almost nobody else offers: the full agent trajectory is published as a HuggingFace dataset — the raw JSONL rollout of the session, timestamped 5 Sep 17:40. Anyone can read what was asked and what the model did. There is also a gameplay video and a documented local run.

It has one star.

Set that against the LoL clip's 344,000 views and you have the cleanest measurement in this article: the project that published its evidence got roughly one three-hundred-thousandth of the attention of the project that published a gif. I have not run it — there is no hosted demo, only a local build — so what I am vouching for is the paper trail, not the gameplay.

6. Two more that are actually live

湖畔 / Lakeside (source) — a Rusty Lake-style point-and-click escape game, eight scenes, Chinese and English, described by its author as made with GPT-6 Astra high. Opens and plays.

NULLSPACE — a survival-horror FPS with a premise built around sound: the building hears you. The site is live and handsome, and it is a site, not a playable game — there is a beta download, which I did not run. Listed because it is real and in development, not because I can vouch for the game.

7. The non-game programs

Almost all the attention went to games, which is a reporting artefact rather than a fact about the model. The GitHub topic gpt-6-astra currently holds 41 repositories, and the useful ones are mostly tools:

Project	What it does
openai-relay-for-cursor	Local relay that translates Cursor's Chat-Completions requests to the Responses API, so a Responses-only model can be used in Cursor's agent with your own key
save-my-astra	Routes cheap work to a smaller model to conserve Astra tokens
atelier	Offline headless pixel-art editor driven by CLI or MCP, described by its author as entirely AI-generated
jiorimetry	Minecraft redstone circuits derived from signal-strength rules rather than transcribed, verified in three tiers

Two of the four exist because Astra is expensive or awkward to access — which is its own comment on the release. Note also that a GitHub topic is self-assigned: appearing under gpt-6-astra is a claim by the author, not a verification by anyone.

8. The counterweight: CAD

The most consequential backlash was not about games. Independent creators posted Astra-generated mechanical parts — a rotating turbocharger split into working components, a camera broken into more than a hundred parts, a car separated into several hundred individually named pieces — and the demos travelled fast.

Engineers who tested it were direct about the problem, as reported on 8 September:

• One part hit 96.1% geometric similarity and was still the wrong part, because a later feature quietly closed an opening the design had to keep.
• On the neuralCAD-Edit benchmark for real edits, frontier models finished 53 percentage points behind humans.
• Robotics builders judged the results as looking impressive but taking more work to fix than to draw correctly from scratch. One seven-hour design run was estimated at a few hundred dollars of usage.
• The headline 95.9% BenchCAD score was achieved in an agentic setting — render the part, measure it, try again — not in single-pass generation.

The pattern is the same one the games show, in a domain where being wrong costs more than a bad review: output that is convincing at a glance and wrong in a way only the domain expert sees.

What the week actually shows

Attention and evidence are anti-correlated here. The ranking by reach is: an unplayable clip (344k), a card game (286k), a shooter with a patch archive (237k), a multiplayer FPS (253k), and a documented one-session build with a published agent trajectory (1 star). The more checkable a project was, the less it travelled.

The good ones are not one-shots. Every project in this list that holds up under inspection has visible iteration: ten timestamped releases, or documented map ports with checksums, or a published session trajectory. The word 딸깍 describes the first prompt, not the artefact.

The bottleneck moved to taste and operations. These builds fail at server capacity, netcode, balance and cheating — not at rendering or code generation. The LAST SIGNAL author is not debugging renderers; he is arguing with a player about whether the final boss should end the run.

Verification is nearly absent, and it is cheap. A hosted link, a repository, a session log. Smash Karts published all three and nobody looked. Until readers start asking, the ranking will keep rewarding the gif.

What this does not show

Reach figures are single readings taken 9 September 2026 and will have moved. They count forum views, which are not comparable to GitHub stars — a repository is a much higher-friction thing to engage with, so the 344,000-to-1 comparison is a statement about visibility, not about relative merit.

Build claims are authors' claims. I verified that each project exists, opened it, and screenshotted it; I did not verify that any of them was made the way its author says, because none published enough to check except Smash Karts, which I read rather than ran. The CAD figures are reported by IBTimes, not measured here.

This is Korean forums plus GitHub plus one English-language report. Reddit's Astra content in this window skews to 3D and Blender demos rather than shipped programs, and X was not searched. Absence from this list is not evidence of absence.

FAQ

What is the most impressive thing an individual has built with GPT-6 Astra?

By verifiable evidence rather than reach: Smash Karts Arena, an 8-player 3D kart battler with server-authoritative simulation, published with its full agent trajectory as a HuggingFace dataset. By playability and polish: Astra Attack, a live 4v4 browser FPS, and LAST SIGNAL, which has shipped ten versioned releases in two days.

Can I play these?

Yes, for four of them, free and with no login: Astra Attack, LAST SIGNAL, Arcane Ascent, 湖畔. All four opened and ran on 9 September 2026.

Are these really made with one prompt?

Almost certainly not, in any case where there is enough information to tell. LAST SIGNAL has a patch archive with ten releases; Astra Attack ships documented asset ports with pinned revisions and checksums; Smash Karts claims one prompt and publishes the session log so you can check. The projects that make the strongest one-click claims are the ones with the least attached.

Why did the League of Legends one go the most viral?

It is a six-second clip of something instantly legible to a large audience, with no link to break the impression. Verification is friction, and friction reduces reach — which is the least comfortable finding in this article.

Is GPT-6 Astra good at CAD?

It scores 95.9% on BenchCAD, but in an agentic setting where it can render, measure and retry. Engineers testing real edits report frontier models 53 points behind humans on neuralCAD-Edit, and one part with 96.1% geometric similarity that was still wrong because a later feature closed an opening it needed. Treat generated parts as sketches, not drawings.

What are people building that isn't a game?

Mostly tooling around the model itself — a relay to use it inside Cursor, a router that offloads cheap work to a smaller model to save tokens — plus a headless pixel-art editor and a Minecraft redstone derivation project. The games get the attention; the tools get the use.

---

← Back to all posts