I Gave Claude 24 Hours to Build GTA 6: How Far Can Autonomous AI Go With External Tools Chained In?

To test how far AI can go working fully autonomously, the creator gave ChatGPT and then Claude 24 hours each to tackle recreating one of the most ambitious games in history, GTA 6. Rather than relying on a single model, the agents were chained to a suite of external AI services -- Tripo for 3D model generation, Nano Banana Pro for textures, ElevenLabs for sound effects and music, and a cinematics tool -- to push the boundaries of what the agents could reach on their own.

Asked for a real GTA-style driving interface, Claude responded in remarkably little time, implementing a driving system complete with acceleration and physics. Being able to hop into a car and actually drive around impressed the creator, who admits he didn't expect it to come out this well.

The next challenge was adding NPCs. Unlike static objects like buildings, NPCs require animated 3D models, making them a notch harder -- and the creator watches to see how the AI handles it.

As development progressed, the project reached a point where almost anything the creator asked for got built. A final cost table broken down by provider showed Claude as the most expensive model of the bunch -- but he judged that a fair trade given it was also the most capable.

The code itself is 'probably a mess,' he admits, partly due to imperfect prompting, and the result is nowhere close to a commercially shippable game. Still, he closes by noting that the sheer level of autonomy AI agents have reached is itself the real story.