OpenAI just dropped GPT 5.6 Sol, Terra and Luna and I decided to test it on a one-shot horror game - single HTML file, enemy AI, inventory, crafting, dynamic lighting, procedural audio. it delivered the whole thing in one shot. genuinely Fable class output. It however also burned 100 % of my daily usage in one prompt. Sol Ultra runs subagents under the hood, so one prompt can burn several runs' worth of tokens. so the real question for 2026: would you rather have a cheaper model you can prompt 50 times, or a monster you can afford to prompt once?

@BuildFastWithAI Looks really cool