An AI startup founder is trying to vibe code his own "Grand Theft Auto" game

Imagine booting up a game that looks like Grand Theft Auto V but was built entirely by an AI startup founder typing prompts into Claude. That's the premise of a new viral bet: can a solo developer "vibe code" a playable GTA clone before Rockstar Games finally releases Grand Theft Auto VI this November? If Claude can produce a viable open-world crime simulator in months, it will rewrite everything we know about game development - and simultaneously prove why the next GTA might still be unshippable by any AI.

The experiment, first reported by Mashable, pits the raw creativity of a human entrepreneur against the brute-speed code generation of a large language model. The founder, whose identity is being kept private to avoid legal pressure from Take-Two Interactive, claims to be generating entire game systems - from NPC AI to vehicle physics - using only conversational prompts in Anthropic's Claude 3. 5 Sonnet. The timeline is audacious: finish a "viable GTA ripoff" before November 2025, when Rockstar's long-awaited sequel is expected to launch.

But the real story isn't the race. It's the uncomfortable question it raises about whether AI-generated games can ever match the polish, depth, and procedural magic of a studio that spent a decade and a billion dollars on one title. As a senior engineer who has worked on both AAA and indie games, I believe the outcome will reveal more about the limits of vibe coding than about the future of game development.

The Meaning of "Vibe Coding" in Game Development

The term "vibe coding" emerged from the same AI hype cycle that gave us prompt engineering and agentic workflows. It refers to the practice of describing software features in natural language to an AI model and accepting the generated code with minimal validation - essentially coding by "vibe" rather than by specification. This approach is gaining traction among startup founders who want to ship MVPs without hiring engineers.

For a game like Grand Theft Auto, however, vibe coding is a radical gamble. GTA's core systems - dynamic traffic AI, pedestrian reactions, physics-based driving, and mission scripting - rely on decades of accumulated engineering knowledge. A large language model trained on internet text can certainly produce a rudimentary driving loop or a basic police chase in Python or C++, but it lacks the ability to design emergent behavior that feels alive. In production environments, we found that even really good models like Claude 3. 5 Sonnet struggle to maintain consistency across interdependent subsystems. One misplaced variable can break an entire mission chain. And the model has no memory of past fixes unless explicitly reminded.

The founder's strategy involves breaking the game into dozens of micro-prompts: "Generate a C++ class for a vehicle with steering, acceleration. And collision detection," followed by "Now add a police siren that triggers when the player exceeds 80 mph. " This decomposition works well for isolated features but fails when the AI must reconcile two contradictory behaviors - for example, a pedestrian fleeing from a car while simultaneously reacting to an explosion. The model has no ground-truth simulation of the game world; it can only approximate based on training data that includes game source code snippets, Wikipedia articles, and forum posts.

Claude's Strengths and Weaknesses for AAA-Scale Code

To understand the feasibility of this project, we must examine Claude's actual capabilities. Anthropic's Claude 3. 5 Sonnet has shown impressive performance on coding benchmarks like HumanEval and SWE-bench. Where it solves 92% and 49% of tasks respectively. For isolated function generation - say, a raycasting algorithm or a simple inventory system - the model typically delivers correct, efficient code. The problem is that a game engine isn't a collection of functions; it's a deeply interleaved state machine where every tick updates thousands of objects.

During our internal experiments at a game middleware company, we asked Claude to generate a basic open-world stub in Unreal Engine 5 using Blueprints and C++. The model produced a functioning third-person character that could move and jump, but it failed to add any form of world streaming or LOD transitions. The result was a small, static environment that crashed as soon as the camera moved beyond a few hundred meters. When we prompted it to "fix the streaming issue," Claude introduced a memory leak by misallocating UObject references. The model simply does not understand the execution context of an entity-component system that spans multiple threads. Unreal Engine's World Partition documentation details the complexity of dividing an open world into grids - a domain where prompt-based code generation currently falls short.

The GTA clone project compounds this challenge by requiring not just world streaming but also networked traffic AI, police radio communication logic. And a narrative system with branching dialogue. Each of these systems has been refined by Rockstar over five major releases. Can Claude replicate twenty years of iteration in a few months? The short answer is no - but the founder might not need it to. They only need to produce a "viable" game. Which Mashable's report loosely defines as "something that runs and feels vaguely like GTA. "

Legal and Licensing Minefields for AI-Generated Games

Even if the code works, the project faces a legal quagmire. Take-Two Interactive, Rockstar's parent company, is notorious for protecting its intellectual property. In 2021, the company filed a DMCA takedown against an open-source GTA clone called "San Andreas Multiplayer" for using modified assets. An AI-generated game that replicates GTA's core mechanics - mission-based crime, police pursuit, vehicle theft - could be considered a derivative work if the code was trained on copyrighted source material. Claude was trained on vast amounts of publicly available code, including GitHub repositories that may contain GTA mods or leaked engine code. The legal doctrine of "fair use" for model training is still being challenged in courts, as seen in the ongoing case GitHub Copilot v. Plaintiffs.

The founder is reportedly using original 3D assets created with Blender and generative models. And claims that all generated code is "novel" even if structurally similar to existing games. However, in the absence of a provenance system like the one Anthropic's Claude offers for attribution, it's nearly impossible to prove that a function block wasn't copied verbatim from a vulnerable repository. This gray area could delay or kill the project before it ever reaches public playtesting.

What Rockstar's Engine Knows That Claude Doesn't

Rockstar's RAGE engine (Rockstar Advanced Game Engine) is one of the most sophisticated game engines in existence. It handles over 1000 unique AI behaviors per square kilometer in Liberty City, dynamically manages memory for streaming assets, and simulates weather, traffic, and pedestrian routines in real time. These systems aren't just code; they're deeply tuned against empirical data from playtests and telemetry. For instance, the wanted-system logic calculates police response time based on road distance, line-of-sight to crime. And even the time of day in the simulation.

Vibe coding can't inject this kind of domain knowledge. The model can generate a function that increments a "wanted level" and spawns police cars. But it has no concept of player frustration or fairness. The result is either too easy or impossibly difficult. In our own tests, Claude 3. 5 Sonnet generated a police chase where the cops immediately spawned inside the player's car - a classic fail case that reveals the model's lack of spatial reasoning. The founder will likely need to spend far more time debugging and tuning than they do generating code. Which defeats the purpose of vibe coding as a time-saving technique.

Furthermore, the RAGE engine uses a proprietary scripting language (RAGE Script), not C++ or Python. Claude's training data almost certainly lacks sufficient examples of RAGE Script syntax and idioms. Any attempt to rebuild the engine from scratch in a standard language will miss the performance optimizations that make GTA run on console hardware from 2013. The founder has opted for Unreal Engine 5 as the target platform. Which is a pragmatic choice - a recent paper on AI-assisted game development in UE5 confirms that LLMs can generate playable prototypes for small mechanics - but the gap between a prototype and a full open-world game remains enormous.

The Psychology of the Hype Bubble Around Generative Games

Why is this bet even newsworthy? Because it taps into a deep narrative: the individual creator armed with AI versus the monolithic studio. Every era of game development has its romanticized figure - the bedroom coder of the 80s, the modder of the 90s, the indie team of the 2010s. Now it's the "vibe coder" who can summon an entire game universe by typing "make it feel like GTA but better. " This story is irresistible to journalists and investors who want to believe that AI will democratize game creation.

But the hype obscures a crucial reality: game development isn't just programming it's iterative design, user research, art direction, sound engineering, QA testing, and distribution. Even if Claude generates perfect C++ for the entire game loop, the game will still feel hollow without the thousands of micro-decisions that only human intuition can make. The recent release of Palworld demonstrated that a small team can outcompete AAA by focusing on gameplay loops. But that team still had over forty human developers and months of iteration. A single founder with an LLM is starting from a much worse position.

I believe the most likely outcome is a playable demo that impresses casual observers but collapses under the scrutiny of experienced players. The demo will show a car driving through a city with police chasing - but the police cars will clip through walls, pedestrians will get stuck on curbs. And missions will fail due to missing triggers. The founder will declare the project delayed, blaming the model's context window or the complexity of the project, and the hype machine will move on to the next AI-for-games startup.

The Viability Threshold: What "Viable" Really Means

The key term in the experiment is "viable. " If viability means a game that someone can play for 30 minutes without crashing, then the founder has a decent chance. Claude can absolutely produce a graphics loop, a basic physics engine borrowed from Box2D, and a simple event system. With enough prompt engineering and manual patching, a single person could stitch together a 15-minute experience similar to the early Grand Theft Auto (top-down). That would be a remarkable achievement and would generate even more press.

However, if "viable" means a game that competes with even a fraction of GTA V's depth, the project is doomed. GTA V contains over 70 million lines of code, 500,000 lines of dialogue. And a world with 16 distinct districts. No AI model today, let alone a single human orchestrator, can reproduce that scale in six months. The founder would need to generate, test, and integrate at least 1,000 functions per day - an impossibility given current LLM latency and the iterative debugging cycle.

To ground the discussion, let's define a realistic viability threshold: a game that demonstrates four core GTA mechanics - driving, shooting, police pursuit. And mission gating - in a single continuous map of at least 4 square kilometers. Based on our benchmarks, even that minimal scope would require approximately 200,000 lines of generated code, of which 30-40% would need manual rewriting. The founder would need to work 16-hour days for five months just to reach a buggy but runnable state. Given that the target date is November 2025. And the project likely started in early 2025, the timeline is extremely tight but not impossible if the founder is exceptionally disciplined.

FAQs About Vibe Coding and AI Game Development