A Taiwanese developer, aicodewithme, has released a free 3D browser game in which players can walk and drive around Taipei. In a Threads post dated September 28, the developer introduced it as a "Taipei GTA" and said they had already spent more than $10,000 on AI tokens. The developer also said they split the work among many AIs while personally putting in 5 to 6 hours a day for about a month.
Looking at how the city of Taipei was turned into a playable game also shows what work remains for humans as more tasks can be handed to AI.
Turning Taipei into a game you can play in the browser
The game can be started from the public site without installing a dedicated app or registering an account. On the screen we checked on October 5, the title was displayed as "TAIPEI RUSH / 臺北狂飆." The site also includes a link identifying aicodewithme's account as the game's only official Threads account. The name used at the time of the post differs from the current title, but it is the same work at the same URL.
When the game starts, a third-person character appears against the backdrop of Ximending and follows an acquaintance named A-Ming. The controls cover walking as well as riding scooters and driving cars. There is also a map that can be opened, and missions can be accepted from an in-game smartphone. Chinese-language signs on buildings and shops along the roads give the game a distinctly Taipei atmosphere.
In 4Gamers' own hands-on play on September 29, the outlet found landmarks such as Taipei 101, Longshan Temple and Shilin Night Market, along with 93 missions according to the in-game display at the time. The story is also rooted in Taipei: the protagonist comes to the city and tries to protect the old neighborhood where their grandmother lives from urban redevelopment. Rather than simply lining up real buildings, the game works the circumstances of the people who live there into its story.
However, it would be a mistake to see it as a simulator that accurately reproduces Taipei's geography. 4Gamers points out that while the game uses real street names, the road layout and the positional relationships between landmarks differ from reality. The fun of visiting familiar places should be judged separately from geographic accuracy.
Over $10,000 in AI costs, and the adjustment work humans handled
The "over $10,000" figure is the developer's own statement in the original post about spending on AI tokens. Tokens are a unit for counting the amount of text and other data an AI processes; here the figure refers to what was paid for using various AI services.
No detailed billing breakdown has been shown, though. It is unclear which models were used and how much, or how much went to code generation versus image and asset creation. The figure should not be taken as total development cost, since it does not include labor or post-launch server and delivery costs.
The development period is explained in more detail in an exclusive interview published by Taiwan's Mirror Media on September 30. The developer worked after finishing their day job, spending about one month at 5 to 6 hours a day. They also said they had "200 AIs" help, from initial research gathering road information through to development.
It is not clear, however, whether "200" refers to 200 different AI models or to a count of execution units such as multiple AI agents or sessions. It can't be read as equivalent to 200 developers, nor as 200 AIs running simultaneously.
The developer also said that adjusting for conflicts between the results produced by multiple AIs was difficult. They did not say what kinds of conflicts arose, but it is clear that a human played the role of pulling together the divided work and ensuring consistency in the end.
If AI-assisted creation is seen as "give instructions and the game completes itself," this workload gets overlooked. The developer's own account includes both handing a large amount of work to AI and continuing to invest a great deal of human time.
No documentation has been released detailing which parts were generated by AI and which were corrected by humans, so it is impossible to calculate how many times more efficient the process was than conventional game development.
Still, a game with missions and means of transport has been released for free and is complete enough for anyone to actually play. It goes beyond showing in-progress footage or demos, reaching a stage where users can try it and judge for themselves.
When considering the effect of using AI, looking at the scale of the finished work together with the amount of human effort put into it is closer to reality than concluding up front that it was "made cheaply."
Instant in the browser, but the device still has to do the processing
In the original post, the developer recommends playing on a PC for higher image quality. Being able to start right away in a browser is not the same as running smoothly on every device.
In 4Gamers' hands-on play, parts of the screen appeared black, there were load waits for scenes at high-quality settings, and audio problems were reported. This reflects the experience as of September 29, and the same issues do not necessarily remain today.
Thinking about how browser 3D games generally work makes the difference easy to understand. According to MDN's WebGL documentation, WebGL is a technology for rendering 3D graphics and more in the browser without additional plugins, and it uses the processing power of the user's device, such as its GPU.
So even though the hassle of installing a dedicated app is eliminated, the computation needed to render buildings, people and vehicles does not disappear. How smoothly the city can be displayed also depends on the performance of the device being used.
The delivery side needs separate processing as well. The game's public URL uses a Vercel domain. According to Vercel's official documentation, a CDN that delivers deployed content from locations close to users is available by default. For delivering shared data such as a game's program files and images to many users, existing delivery infrastructure like this can be used. An individual developer does not need to build a delivery network from scratch.
But delivering static assets to many users is a different problem from synchronizing multiple players' positions and actions in real time. The latter requires managing the constantly changing in-game state on the server side and continuously sharing it with multiple devices.
The mere fact that the game is hosted on Vercel does not reveal how it implements multiplayer or how many simultaneous connections it can handle.
In fact, in response to a reaction to the launch announcement saying "I can't see other players," the developer replied on September 28 in a reply that the feature is under development. They also explained that there were many users and the server was overloaded.
This describes the situation at the time and does not confirm the current state of implementation. Still, the number of people who can access the game at the same time should be distinguished from the number who can play together in the same virtual space while seeing one another.
"1.2 million concurrent players": the definition of the number needs checking
In an article on October 2, Taiwan's TVBS reported that 1.2 million people played online at the same time within three days of release.
The article, however, does not show an analytics screen or how the figure was tallied. It is also impossible to tell whether it is the number of concurrent users at a single moment or the number of users over a period, or how repeat visits by the same person were handled. It is a very large number, but it cannot be directly compared with typical game concurrency records whose measurement conditions are public.
There is also a discrepancy between sources on the release date. The developer announced the free release in a post on September 28, while TVBS's article says it was released on October 1.
| Source | Date / description | What the comparison shows |
|---|---|---|
| Developer's announcement of the free release | September 28, 2026, Japan/Taiwan time. Says "you can play for free right now" | The developer themselves had announced the free release by at least this point |
| TVBS's original reporting | Published October 2, 2026. States "released October 1" and "1.2 million online at the same time on day 3" | The release date does not match the September 28 announcement, and the measurement conditions for 1.2 million are not shown |
The original post's timestamp is 17:21 UTC on September 27, which is 2:21 on September 28 in Japan and 1:21 the same day in Taiwan. Even allowing for time zones, the free release was announced before October 1.
However, the time of the Threads post does not necessarily match the moment the game was first made public, and the basis for TVBS's "released October 1" is not clear.
This discrepancy alone is not grounds to declare the "1.2 million" figure wrong. Nor is there any basis for rereading it as cumulative users or views of social media posts. For now, it is reasonable to treat it as a usage statistic reported by TVBS whose measurement method and definition cannot be verified.
As the map grows, the question is whether people will want to keep playing
In the Mirror Media interview, the developer said they hope the game will help people get to know Taiwan, and that both people living in Taiwan and those living away will remember time spent with their families. This explains why they chose Taipei as the subject.
Rather than simply aiming for a vast map like a blockbuster game, they are trying to express in game form the streets they know and their memories of life in the area.
The original post said the map is still being expanded and invited users to leave comments about what they would like added. With a free format that can be played just by opening a link, interested people can experience a world modeled on Taipei without any complicated setup. The developer's intent in choosing the subject combines with an easy-to-try release method.
The appeal of this work cannot be explained by the numbers on how much AI was used alone.
Going forward, challenges will multiply: how to connect missions each time the map is expanded, and whether movement on foot and in vehicles can run stably on different devices. Responding to bug reports from users and bearing the costs needed for continued delivery are also part of game development.
Whether a work that AI helped bring to release in a short time can grow from something tried once into a game people want to play repeatedly is the question. What will be tested is whether the developer can keep improving not just the size of the map but also playability and the completeness of the content.