I Ran GPT-6 for 3 Days Non-Stop and It Created This…

Summary
Stefan (from Stefan 3D AI) evaluates the capabilities and limitations of OpenAI’s GPT-6 (specifically GPT-6 Astra) for end-to-end game development under a strict “prompting-only” condition. Across multi-day sessions consuming billions of tokens, he documents the creation of two distinct projects: a first-person zombie horde shooter in Unity built iteratively from a 2D prototype to realistic 3D, and a Miami-style open-world GTA 6 clone developed in Unreal Engine 5 by his friend.

What is shown

• [00:16] Setting the test parameters: the developer must guide the process exclusively through prompt engineering and terminal/CLI workflows.
• [01:28] Rapid generation of an initial 2D top-down arena shooter prototype (Dead / Signal) in Unity using Codex CLI, Unity CLI, and MCP servers (including official Unity plugins and Anklebreaker Unity MCP).
• [02:22] Upgrading the 2D layout into a playable low-poly 3D first-person arena shooter complete with UI, waves, and basic procedural logic.
• [03:15] Iterating into a realistic horde shooter environment (Dead Signal: Redline) with custom lighting, atmospheric effects, and weapon mechanics.
• [04:39] Utilizing the Higgsfield Blender plugin and MCP to coordinate multi-modal asset generation (Tripo, Meshy, Seedance, Nano Banana Pro) for zombie character models, texturing, rigging, and animations.
• [06:13] Iterative debugging of first-person weapon reload and firing animations for an assault rifle, double-barrel shotgun, rocket launcher, and minigun.
• [07:41] Demonstration of the final playable Unity horde shooter wave-defense gameplay.
• [08:50] Inspection of telemetry and production costs: 1.29 billion tokens, 302 generation requests (103 images, 58 3D models, 35 animation jobs, 54 audio tracks), totaling approximately $40–$85 in asset generation costs.
• [10:06] UE5 GTA 6 recreation workflow: initial low-poly driving prototype evolving into a textured Miami Beach beachfront with day/night cycles and MetaHumans.
• [12:05] Solving complex character-vehicle interaction animations by feeding real video footage of a person entering a car into GPT-6 via the GVHMR (Ground-to-Human Motion Recovery) library.
• [12:47] Generation and placement of multi-story modular Miami architecture, walk-in interior spaces (café/bakery, surf shop), and traffic AI.
• [13:57] Final UE5 gameplay showcase and project audit: 2.53 billion tokens consumed over 4+ days, generating 247 voice lines, 5 animation pipelines, and 8 fully populated buildings.

Claims & numbers

• The presenter spent 1.29 billion total tokens (8.50M generated output; 46 hours 25 minutes active processing; 107 hours calendar window) on the Unity shooter project.
• Asset creation for the Unity game yielded 103 images via Nano Banana Pro, 58 3D models via Tripo H3.1 and Meshy V7, 35 rigging/animation jobs via Seedance 2.5 and Higgsfield, and 54 audio files via Seed Audio and Mirelo.
• The presenter estimates the asset generation costs for the Unity project at $40 to $85 (max ~$100).
• The UE5 GTA project consumed 2.53 billion tokens over more than four continuous days of prompting.
• The UE5 project generated 247 unique voice lines (costing ~$7–$25 across 203 resident lines and 44 police/driver lines), 22 ambient sound sources, 5 radio tracks across 2 stations, 51 preserved environment/prop models, and 16 PBR material sets, totaling an estimated generation cost between $40 and $115.
• The presenter claims that relying solely on AI prompting is dramatically slower and less efficient than combining developer expertise with targeted AI assistance.

Notable quotes

• [01:13] “I’m going to need to move into iterations because that one-shot approach, it’s a dead end.”
• [08:08] “When you combine AI tools with your real skills, then the real magic happens.”
• [14:41] “I definitely think there’s a better way to spend 2 billion tokens and four days of your life.”

Assessment
This is a genuine developer retrospective and technical demo evaluating GPT-6 Astra’s current automated coding and multi-agent pipeline limits in Unity and Unreal Engine 5. The presenter shows actual telemetry logs, Git/CLI terminals, and raw unpolished intermediate glitches (such as broken rigging and disjointed meshes), demonstrating real-world workflow strengths and bottlenecks without exaggerated marketing hype.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.