Is that input or output tokens or both? That number sounds quite extreme. Maybe they include input tokens from deep research? That could be tens of thousands of input tokens into a cheap model per task, for example.