⚡ Overview

We tested three frontier AI models on a brutal coding challenge: generate a complete, movie-quality Three.js solar system simulator in a single HTML file — all procedural textures, custom shaders, bloom post-processing, and interactive controls, zero external assets.

🎯 The Challenge
Same prompt, three models, head-to-head. The prompt specified 15+ celestial bodies with exact shader requirements (multi-layer FBM noise, vortex noise, cellular noise), 3 glow layers on the sun, Jupiter's Great Red Spot with spiral edges, Saturn's Cassini division, Earth's city lights on the dark side, and more. No shortcuts allowed.

🤖 Models Tested

ModelToolTask 1: Solar SystemTask 2: City Renderer
GPT-5.6 SolCodex11 min11 + 33 min (2 rounds)
Kimi K3Kimi Code45 min3+ hours, round 2 abandoned
Qwen3.8MaxDirect<10 min (2 rounds)
⚠️ Methodology Note
This is not a formal benchmark — it's a real-world stress test by GitHub user OBdangshang07. Same prompt, same expectations, measured wall-clock time from tool invocation to complete output. Quality judged by visual fidelity and code completeness.

🌌 Task 1: Solar System Simulator

The prompt demanded a cinematic-grade solar system with 15+ celestial bodies, each requiring custom ShaderMaterial with multi-layer noise. Key requirements:

  • Sun: Dynamic plasma texture with 3 glow layers, sunspots, solar wind particles
  • Jupiter: Multi-layer stretched noise bands, Great Red Spot with spiral edges, 3+ smaller vortices
  • Saturn: Multi-layer translucent rings with Cassini division, particle scatter via shader discard
  • Earth: Ocean specular highlights, city lights on dark side, Fresnel atmosphere shell
  • Moons: 4 Galilean moons + 4 Saturnian moons, each with unique procedural textures

Results

AspectGPT-5.6 Sol (11 min)Kimi K3 (45 min)
Visual fidelity⭐⭐⭐⭐⭐ Excellent shaders, Jupiter bands look organic⭐⭐⭐⭐ Good but some bodies use simpler noise
Code completenessAll 15+ bodies, all moons includedAll bodies present, some moons simplified
InteractivitySmooth orbit controls, info panels, time warpSimilar controls, slightly less polished UI
Post-processingUnrealBloomPass well-tunedBloom present but less refined
StabilityStable at 60fpsOccasional frame drops on lower GPUs

🏙️ Task 2: City Renderer

A harder challenge: generate a real-time metropolitan renderer with 50+ adjustable parameters, procedural infinite city, dynamic day/night cycle, atmospheric scattering, and dense 3D building clusters with unique textures.

Results

ModelRound 1Round 2 (Refinement)Building DensityTexture Quality
Qwen3.8Max<10 min<10 min⭐⭐⭐⭐⭐ Dense, varied⭐⭐⭐⭐ Good procedural
GPT-5.6 Sol11 min33 min⭐⭐⭐⭐ Dense⭐⭐⭐⭐⭐ Excellent detail
Kimi K33+ hours❌ Abandoned⭐⭐⭐ Moderate⭐⭐⭐ Decent
🚨 Kimi K3 Struggled
Kimi K3 took over 3 hours for the first round and couldn't complete the second. The city renderer task exposed K3's weakness in generating complex, performance-sensitive 3D code with many interdependent systems (roads, traffic, lighting, building generation).

📊 Key Findings

🏎️ Speed

Qwen3.8Max is the speed king — both rounds under 10 minutes total. GPT-5.6 Sol is consistent (~11 min per task). Kimi K3 is 3-4x slower on complex tasks.

🎨 Quality

GPT-5.6 Sol produces the best visual quality — its shaders are more refined, noise layers more organic, and post-processing better tuned. Qwen3.8Max trades some visual polish for raw speed. Kimi K3's output is functional but less cinematic.

🔧 Reliability

GPT-5.6 Sol is the most reliable — it completed both tasks fully. Qwen3.8Max is fast but may need a refinement round. Kimi K3 can timeout on complex multi-system tasks.

🎮 Live Demos — Try Them Yourself

Below are the actual outputs from each model. Open them in your browser and compare the visual quality, interactivity, and performance:

☀️ Solar System Simulator

ModelGeneration TimeDemo
GPT-5.6 Sol11 min🚀 Launch Demo →
Kimi K345 min🚀 Launch Demo →

🌃 City Renderer

ModelGeneration TimeDemo
GPT-5.6 Sol11 + 33 min🚀 Launch Demo →
Kimi K33+ hours🚀 Launch Demo →
Qwen3.8Max<10 min × 2🚀 Launch Demo →

🎯 What This Means for Developers

💡 Model Selection Guide
  • Need the best visual quality? → GPT-5.6 Sol (Codex). Best shaders, most polished output.
  • Need speed? → Qwen3.8Max. Blazing fast, good enough quality for prototyping.
  • Complex multi-system 3D? → Avoid Kimi K3 for now. It excels at simpler tasks but struggles with interconnected systems.
  • Cost-sensitive? → Qwen3.8Max delivers the best time-to-quality ratio.

The gap between models is not just speed — it's about the complexity ceiling. GPT-5.6 Sol can handle prompts with 15+ interdependent shader systems. Qwen3.8Max is fast but may need refinement passes. Kimi K3 has a lower ceiling on complex 3D tasks.

For single-file HTML generation — the emerging "vibe coding" benchmark — GPT-5.6 Sol currently leads in quality, while Qwen3.8Max leads in throughput. Choose based on your priority.

📎 Source

All code, prompts, and test results are open source: github.com/OBdangshang07/AI_project

Test conducted by GitHub user OBdangshang07. Demos reproduced with original HTML output files.