Ziwen Xu

An AI startup founder is attempting to build a playable version of Grand Theft Auto VI using Claude Max before Rockstar’s official September release, highlighting both the accelerating capabilities of frontier AI coding models and the widening gap between what AI can theoretically generate versus what constitutes a complete, polished game product. For institutional investors tracking AI infrastructure adoption, the experiment reveals real constraints on Claude’s practical utility for large-scale game development despite sustained hype around AI’s coding prowess.

• Ziwen Xu, founder of AI agent startup Hypercho, launched a GitHub-hosted GTA 6 recreation on June 10, paying for Claude Max 20x access specifically for the project.
• Within 24 hours, Xu consumed 33 percent of his weekly Claude Max token allowance, exposing cost and rate-limit barriers to sustained large-scale AI-driven development work.
• The project’s early failures, generating Los Angeles architecture for a Florida-set game, demonstrate that frontier language models still lack grounded reasoning for coherent game world design at scale.

• June 10 Official launch date of Xu’s vibe-coded GTA 6 project on public GitHub repository
• 33% Weekly Claude Max token budget consumed in first 24 hours of development work
• Sept 2025 Stated release window for Rockstar’s actual GTA 6, Xu’s stated competitive deadline

Ziwen Xu, the founder of AI agent startup Hypercho, announced on June 10 that he would attempt to build a playable version of Grand Theft Auto VI before Rockstar Games releases the official title this September.

In a tweet pairing a video of a humanoid character navigating obstacles across a procedurally generated landscape, Xu declared his ambition with a mixture of confidence and self-aware irony: “The goal: beat the real GTA 6 to launch.

Ambitious, probably stupid, doing it anyway.” The project relies entirely on Claude Max 20x, Anthropic’s highest-tier large language model tier, which Xu explicitly upgraded to specifically for the task.

Behind the apparent jest lies a genuine technical experiment, one that exposes both the genuine capabilities and hard limitations of frontier AI coding systems when tasked with real-world, capital-intensive software development.

Xu’s Claude Max subscription exhausts token limits within hours of development start

The experiment turned instructive almost immediately. Within the first 24 hours, Xu revealed that his work had already consumed 33 percent of his weekly Claude Max token allowance.

This constraint proved significant: token-per-week limits are Anthropic’s mechanism for managing compute costs and preventing API exhaustion at scale, and hitting a third of that budget in a single day suggested that even basic game world scaffolding, terrain generation, asset placement, character controllers, basic physics integration, requires sustained, expensive LLM calls that exceed what most individual developers or small studios can absorb.

That cost structure raises a harder question for institutional investors evaluating AI infrastructure moats. If a single developer building a non-commercial proof-of-concept burns through token allocation at this rate, what does actual production game development look like when accounting for iteration cycles, debugging, and the kind of collaborative back-and-forth that real studios require?

The weekly cap, meant as a safeguard, became a practical ceiling on project ambition. Xu noted the ticking clock explicitly, flagging that his remaining budget would not sustain the pace required to ship anything coherent before Rockstar’s September launch window.

Downtown Los Angeles skyscrapers appear in supposed Florida-set game world

The technical failures emerged almost as quickly as the token constraints. On June 11, Xu posted a follow-up noting that Claude had generated detailed architecture appropriate for downtown Los Angeles, clean high-rises, urban density, contemporary commercial design, for a game world supposed to be set in Florida, with its distinct palm-lined streets, art deco buildings, and subtropical sprawl.

The model possessed sufficient architectural knowledge to generate convincing urban environments, but lacked the grounded reasoning to apply setting-specific constraints. It had no internal map of where Miami or Tampa or Jacksonville differ architecturally from LA, and no mechanism to prevent hallucinated geographies that broke the intended game’s coherence.

The video accompanying the complaint showed a humanoid character dressed in a stereotypical burglar outfit traversing a sunset-bathed cityscape that looked, as Xu himself described it, like “the background features in an old FPS that you noclipped into”, detailed enough from a distance to read as a game world, but lacking spatial integrity, navigation logic, or the kind of environmental storytelling that defines successful open-world design.

This distinction matters to investors evaluating AI’s readiness for production creative work: generating assets or code snippets is fundamentally different from orchestrating a coherent, rule-respecting system at scale. Claude could produce individual building models or Python classes, but it could not maintain a consistent, internally logical game world across multiple generation passes.

Matt Shumer’s Fable model proposal sparked Xu’s competitive challenge

The experiment’s origins lie in a proposal from Matt Shumer, an AI entrepreneur and prominent evangelist for Anthropic’s capabilities.

Shumer had publicly suggested that vibe coders should coordinate to use Anthropic’s Fable model, a specialized LLM designed for long-horizon coding tasks, to build a “GTA-VI-caliber open-world game with a quality and scope surpassing what is shown in the initial trailers.” Fable represents a deliberate downgrade from Mytho, a more powerful Anthropic research model that the company refused to release publicly, citing risk that its coding capabilities could enable zero-day exploits against operating systems and web browsers.

Shumer’s provocation, framed as an exciting opportunity for AI-driven creative collaboration, triggered Xu’s public commitment.

The proposal itself signals how venture and AI founder circles now frame model capabilities: not as tools with defined, tested scopes, but as frontier technologies whose limits are genuinely unknown and whose risks scale with deployment. That framing attracts both legitimate technical talent and performative hype-cycling.

Xu’s project exists somewhere in that spectrum, a genuine technical probe into Claude’s practical constraints, partly a commentary on GTA 6’s decade-plus development cycle and repeated delays, and partly engagement-driven content from a startup founder with brand incentive to stay visible in AI discourse.

What distinguishes Xu’s experiment from purely marketing-oriented “AI can do anything” claims is its public documentation. He posted code to GitHub, shared videos of actual failures, and disclosed the hard numerical limits (token budgets, time-to-failure metrics) that constrain the work.

This transparency makes it possible to extract genuine signal about where frontier models’ practical utility begins and ends when applied to real creative engineering problems.

Institutional investors face clearer reckoning on AI coding model production readiness

For institutional investors evaluating AI infrastructure companies and their claims about developer productivity, Xu’s project provides a useful real-world data point.

The messaging around Claude Max, Fable, and other frontier coding models often emphasizes their theoretical capabilities, ability to reason through complex multi-file systems, generate valid code at scale, maintain context across long sessions.

The project’s early results suggest those capabilities are real but bounded by economics, logical consistency, and task-specific reasoning that models do not yet possess reliably.

The model can generate code and assets quickly, but at cost structures that make iteration expensive and at coherence levels that require substantial human supervision and redesign.

A professional game studio would not use Claude Max as a primary development tool under these constraints; they would use it as a narrow accelerator for specific, well-scoped tasks like shader generation or boilerplate API integration. That is a real value proposition, but it is narrower than venture narrative typically frames.

Xu has committed to posting regular progress updates and maintaining the GitHub repository public. The concrete next marker is whether he produces any playable vertical slice, a discrete, complete game sequence (a robbery, a chase, a story mission) that actually runs and functions, before Rockstar’s September release. The outcome will clarify whether AI-powered game development remains a technical stunt or whether models like Claude Max can genuinely accelerate complex creative engineering at scale. The token budget constraint will force Xu to either find external funding to sustain token costs, optimize his prompts drastically, or abandon the project within weeks, each outcome tells investors something different about frontier model economics and real-