Everyone talks about how fast AI design tools are. Nobody mentions the meter running underneath.

Anthropic recently overhauled Claude Design, adding design system imports and a proper round trip with Claude Code. Buried in the announcement was the actual reason for the update. Early users had been burning through most of their weekly Pro token allowance in under thirty minutes, on a single prototype.

That is not a rounding error. That is the real cost structure of AI-assisted design, and almost nobody prices client work around it.

I have watched designers quote a project the old way, then eat the AI tooling cost themselves because they never separated "time I save" from "compute I spend." Speed and cost are not the same line item, and treating them as one is how a fast job turns into a break-even job.

Before your next AI-heavy sprint, actually track what one serious iteration costs in tokens, not just minutes. Then decide if that's a client line item or your own margin.