I gave the same galaxy shooter prompt to GPT-6 Astra, Opus 5.5 and Kimi K3.1 Pro. Same steps, same order, three browser tabs. -> Kimi's build is the one I kept playing, and it cost $0.31. One rule for all three: never ask for "A space shooter". One system per prompt, play it, then the next. Step 1 -> A canvas, a starfield, a ship that moves on arrows and touch. No enemies yet. Step 2 -> Auto fire and one enemy type that dives. Play it until the dodge feels right. Step 3 -> Waves. Each one tighter, with a "WAVE 2" banner so you know the game noticed. Step 4 -> Power-ups: spread shot, shield, score bonus, each dropped from a kill. Step 5 -> Score, best score, lives, a pause key, a game over screen that restarts on space. Step 6 -> Each model reviews its own build and fixes what it finds before I touch anything. What came back: > GPT-6 Astra: $2.55, a ship, enemies, levels. Spread shot never landed. > Opus 5.5: $1.93, clean levels and a proper HUD, no pause, no power-ups. > Kimi K3.1 Pro: $0.31, waves, spread shot, pause, best score, game over, all working on the first run. The split: > HTML canvas: the whole game, zero libraries. > The models: every line of code. > Me: play, compare, keep the one that feels like a game. Speed and cost decided this. Kimi K3.1 Pro on the high-speed version returned each step in seconds and landed at $0.31 total, about 8x cheaper than Astra and 6x cheaper than Opus for the most complete build of the three. Commit after every step that works. A bad prompt costs seconds, not an evening.