Why build a custom fine-tuned LLM instead of using flagship models like Claude, given token costs are low?

The presenter's stated reason was that token costs on flagship models were too expensive at scale. When pressed on whether cost was really the driver (since a single chat query might cost only a couple cents), the presenter said one query cost about 1 cent on their local fine-tuned model versus an estimated 20 cents if run on public cloud flagship models -- a meaningful difference at volume.