My recommendation is to think in cost per task and tokens consumed. Metered use makes each call visible and encourages
restraint; it does not enforce a cap by itself. Set budgets and application limits, and decide which tasks justify a
model call. Local inference also consumes hardware, electricity, and time.