This guy is trying to vibe code GTA 6 before its official release

At a glance:

• Ziwen Xu, founder of Hyperecho, announced a daily AI‑driven attempt to recreate GTA VI before its November launch.
• The project relies on Anthropic’s Claude Fable 5 (upgraded to Claude Max 20x) and is tracked publicly on GitHub.
• Early demos show a blue oval, then a rudimentary urban scene with NPCs, cars and weapons, but the effort has already burned 33% of Xu’s weekly AI quota.

What happened

Ziwen Xu, a 25‑year‑old AI startup founder, posted on X on Wednesday that he had begun “Day 1 of building GTA 6” using Anthropic’s Claude Max 20x model. He shared a short video of a 3D blue oval bouncing among gray blocks, marking the first visual output of his so‑called “vibe coding” experiment. Within hours he structured the project, pushed the code to a public GitHub repository, and pledged to update daily until the real game releases in November.

The effort is not a lone hobby; Xu is the CEO of Hyperecho, a company that helps enterprises deploy “AI employees.” He has been devoting several hours each day to the venture, even after noting that he burned 33 % of his 20× weekly Claude usage in a single day. By day two, a more recognizable character was shown running through an urban landscape—though Xu admitted the environment mistakenly resembled downtown Los Angeles instead of the intended Florida setting.

How vibe coding works

“Vibe coding” describes a workflow where developers issue natural‑language prompts to advanced generative AI agents, which then generate, test, and debug code autonomously. The term was popularized by OpenAI co‑founder Andrej Karpathy, who later called the practice “net unhelpful,” while other high‑profile founders such as Jack Dorsey have embraced it for rapid prototyping. In Xu’s case, the prompt chain began with a request to create a GTA‑caliber open‑world game, and the AI responded by assembling assets, scripting NPC behavior, and even generating simple weapon mechanics.

Anthropic’s Claude Fable 5, a safety‑tuned variant of the company’s Mythos series, is the engine behind the project. Xu upgraded to Claude Max 20x to gain higher token limits and faster response times, allowing the model to handle the massive codebase required for an open‑world environment. The AI iteratively refines the code: after each generation, Xu reviews the output, fixes errors, and feeds the corrected version back into the model for further development.

Progress and challenges so far

By the third day, the repository showcased NPCs walking along streets, cars navigating roads, and rudimentary weapon handling. However, the project faces significant technical hurdles. The AI still misplaces geographic cues (e.g., LA skyscrapers in a Florida map) and struggles with performance optimization required for a seamless open‑world experience. Moreover, the consumption of AI credits is steep; Xu’s 20× weekly quota is a finite resource, and each large‑scale generation cycle can exhaust a sizable portion of it.

Another challenge is the legal and ethical landscape. Rockstar Games, the studio behind GTA VI, has not commented on the initiative, but recreating a proprietary franchise raises copyright concerns. While the code is generated from prompts rather than direct asset copying, the similarity in scope and design could attract scrutiny if the project ever moves beyond a proof‑of‑concept.

What this means for AI‑driven game development

Xu’s experiment highlights both the promise and the current limits of AI‑assisted game creation. On one hand, the ability to produce functional prototypes in days rather than months could democratize indie development and accelerate concept validation. On the other hand, the massive compute costs, token limits, and quality gaps underscore that human expertise remains essential for polishing, balancing, and optimizing large‑scale games.

If the community‑funded “Fable run” idea suggested by Matt Shumer gains traction, we may see collaborative AI‑generated game jams where multiple developers pool resources to push AI models toward producing commercially viable titles. Until then, Xu’s daily updates serve as a live case study of how far current generative models can stretch when tasked with recreating a blockbuster‑level open‑world experience.