Same immediate thought: the free option I provide on my production site is a model that runs on 2xA40. That's 96GB of VRAM for 78 cents an hour serving at least 4 or 5 concurrent requests at any given time.

O3 Mini is probably not a very large model and OpenAI has layers upon layers of efficiencies, so they must be making an absolute killing charging 3.3 cents for a few seconds of compute