I Raced Every Gemini Model inside Antigravity !! (Winner was unexpected. )

Which Gemini model actually understands a normal human prompt the best? I gave four Gemini models the exact same vague prompt and asked each of them to build the same game — a strange blend of GTA Vice City and Minecraft. The models tested were Gemini 3.6 Flash, Gemini 3.7 Flash, Gemini 3.8 Flash, and Gemini 3.1 Pro. I deliberately did not use AI to refine or optimize my prompt. I wanted to test these models the way most people actually use AI: you have an idea in your head, explain it imperfectly, and expect the model to understand what you mean. This isn't meant to be a scientific benchmark. It's a practical experiment to see how differently these models interpret the same request, make decisions, write code, and turn an unclear idea into an actual working game. And more importantly: does the newest or most powerful model necessarily give you the best result for every type of task? Watch till the end to see which Gemini model handled the challenge best. Models tested: Gemini 3.6 Flash Gemini 3.7 Flash Gemini 3.8 Flash Gemini 3.1 Pro #Gemini #GoogleGemini #AICoding