原帖内容

This comparison got way more interesting than I expected I gave grok 4.6 build, gpt-5.6 sol, gemini 3.7 flash, and glm-5.3 flash the exact same prompt + image ref: build an interactive arcade water sim with sharks in three.js all 4 cooked up completely different outputs: ❌ gemini 3.7 (biggest L): output runs fine and tap works, but it completely ignored the reference image aesthetic and built it in a totally different style lol 🌊 grok 4.6: reflections look like a high-budget render and beach matches the ref. sharks look a bit scuffed tho, and ripple physics are way too aggressive 💀 glm-5.3 flash: water looks simple but realistic, sharks are decent, but tap is buggy, and it casually melted 1.5M tokens lmao 🔥 gpt-5.6 sol: by far the closest to the image ref and prompt. tap is mid, but the rest is straight fire (not a pure one-shot run ngl) zero perfect models exist... each has its own pros and cons, and anyone claiming ai can reliably one-shot complex stuff rn is on pure copium