OpenAI just launched GPT-6.2 Astra and Sol, and I decided to test the new Astra with a deceptively simple 3D challenge. Build a realistic glass light bulb hanging from a twisted cable, entirely in WebGL, with realistic materials, lighting, and interactive physics. The result? Surprisingly impressive. And you can play with it yourself. DRAG THE LIGHT BULB AND LET IT SWING: https://niti-svetovaya-studiya.yunitskiy.chatgpt.site/ WHAT DID GPT-6.2 ASTRA BUILD? Using Codex, I asked Astra to create the entire scene from scratch through code. The agent constructed: • Procedural 3D geometry: generated the glass bulb and metal socket by rotating custom profiles around an axis. • Detailed light bulb components: modeled the screw threads and incandescent filament separately. • Glass shader: implemented reflections, highlights, and simulated light refraction. • Interactive cable physics: created a twisted hanging wire with connected segments, gravity, and motion damping. • Real-time interaction: grab the bulb with your mouse, move it around, and release it. • Dynamic movement: the bulb swings naturally before gradually coming to rest. • WebGL rendering: everything runs interactively in your browser. WHY IS THIS AN INTERESTING AI BENCHMARK? A light bulb looks simple, but it's a surprisingly demanding 3D challenge. You need transparent glass, metallic materials, realistic lighting, complex geometry, and a convincing physical response when the object moves. It's a useful way to evaluate whether an AI coding model can combine visual design, shader programming, geometry, and physics in a single interactive experience. And that's exactly what I wanted to test. GPT-6.2 ASTRA VS OTHER MODELS In OpenAI's launch announcement, GPT-6.2 Astra took first place on Artificial Analysis's AA-3D benchmark, where human judges blindly compare interactive 3D experiences created from identical prompts. The reported results: • GPT-6.2 Astra: 1318 Elo • Claude Fable 5.2: 1271 Elo • GPT-6.2 Sol: 1259 Elo Astra is designed for complex projects, while Sol targets faster and more economical everyday work. Of course, one impressive WebGL demo doesn't prove that a model can build production-ready physics simulations or physically accurate optical materials. But the progress in AI-generated interactive 3D is difficult to ignore. And I think I've found a new way to compare AI coding models. THE LIGHT BULB TEST. Give different models the same brief: build a realistic hanging light bulb, with glass shaders, a detailed filament, and interactive cable physics. Then compare the geometry, visual quality, performance, and physical behavior. TRY THE INTERACTIVE DEMO: https://niti-svetovaya-studiya.yunitskiy.chatgpt.site/ Official GPT-6.2 announcement: https://openai.com/index/introducing-gpt-6-2-astra-sol/ AA-3D benchmark: https://artificialanalysis.ai/evaluations/aa-3d Which AI model should take the Light Bulb Test next? Claude Fable, Gemini Carbon, or GPT-6.2 Sol? #GPT62Astra #WebGL #AICoding