A solo developer has been trying to build an AI made version of GTA 6, but the project quickly showed both the promise and limits of AI assisted game development. The experiment produced fast results, including NPCs, traffic, reflections, a phone system, and a playable open world setup, but the AI agents also made major mistakes, including building the wrong city.
The project, called GT Caliber, was started by Ziwen Xu, founder of AI agent startup Hyperecho. Development reportedly began on June 10 using the Godot engine, with AI agents helping create systems and respond to community feature requests.
The results were impressive in speed, but also messy. By the second day, the AI agents had started recreating Los Angeles instead of Miami, even though GTA 6 is set in Vice City, which is inspired by Miami.
The AI project moved quickly at first
The early progress was surprisingly fast. Within three days, the AI agents had added NPCs walking around, cars on roads, shadows, reflections, and several requested features from the community.
That kind of speed is what makes AI game development experiments interesting. A solo creator can use agents to build prototypes much faster than traditional manual development.
| Project detail | What happened |
|---|---|
| Developer | Ziwen Xu |
| Project name | GT Caliber |
| Started | June 10 |
| First engine | Godot |
| Later engine | Unreal Engine 5 |
| Major mistake | AI built Los Angeles instead of Miami |
| Added systems | NPCs, cars, reflections, phone, maps, contacts, wallet |
| Later problem | Unreal Engine 5.8 slowed the build |
The project showed how AI can help generate pieces of a game quickly, but it also showed that speed does not guarantee creative direction or accuracy.
The wrong city became the funniest mistake
The biggest early mistake was the city itself. GTA 6 is built around Vice City, Rockstar’s fictional version of Miami. The AI agents, however, started building something closer to Los Angeles.
That kind of mistake explains why AI still needs strong human direction. The agents can generate assets and systems, but they may misunderstand the creative brief or follow the wrong pattern if the instructions are not precise enough.
For a game like GTA 6, the city is not a background detail. It is the identity of the whole experience. Building the wrong city means the AI missed one of the most important parts of the project.
Switching engines created new problems
The project also ran into technical problems. On Day 7, Xu moved the project from Godot to Unreal Engine 5. On Day 8, his Mac reportedly died because every AI agent opened the game at once to take screenshots.
Even with those issues, development continued. The NPC walking animations were lost at one point, but the AI agents created an in game phone with maps, contacts, and a wallet.
That mix of progress and breakage is typical of early AI assisted development. The tools can create new systems quickly, but they can also introduce problems just as quickly.
Unreal Engine 5.8 did not help as expected
Xu later tried moving the project to Unreal Engine 5.8 to use its Model Context Protocol support, but that did not improve the workflow. According to the developer, the new system slowed the build down instead of making it better.
That is another important lesson. New tools are not always better for every project. Developers often assume the newest engine version or newest AI feature will improve everything, but real production work depends on stability, speed, and predictable iteration.
In this case, chasing the latest version appears to have made the loop worse.
The experiment shows what AI can and cannot do
The GT Caliber project is interesting because it does not simply prove that AI is useless or magical. It shows both sides at once.
AI agents can help a solo developer build a rough open world prototype faster than would normally be possible. They can create systems, generate content, test ideas, and respond to feature requests quickly.
But they do not replace direction, taste, design judgment, or deep technical control. They can make the wrong city, break animations, overload a machine, and slow a project down with unnecessary changes.
GTA 6 remains far beyond AI’s current reach
The experiment also proves why a real GTA 6 scale game cannot be built by AI alone today. Rockstar’s games rely on writing, world design, animation, mission structure, satire, art direction, physics, audio, performance tuning, and years of coordinated work across huge teams.
AI can imitate pieces of that process, but it cannot yet produce the full creative and technical quality of a Rockstar open world on its own.
That is close to what Take Two CEO Strauss Zelnick has said before. Even if AI could help build large parts of a game, creativity remains the missing piece.
A funny but useful look at AI game development
The AI made GTA 6 experiment is entertaining because the results are both impressive and chaotic. A solo developer built a rough open world with AI agents in days, but the project also got confused, broke parts of itself, changed engines, and struggled with newer tools.
That makes it a useful snapshot of where AI game development stands right now. It can speed up prototyping, but it still needs strong human control.
GT Caliber is not a real GTA 6 replacement. It is a fast, strange, and funny example of what happens when AI agents are asked to build something far larger than their current abilities.
Discussion (0)
Be the first to comment.