One task used 30% of my weekly usage.
I used to run this almost around the clock on research and projects and not come close to the cap because the tool calls run as programs and only the result comes back into context. Scripts. Applications I built for the work. That is how I stayed off the meter.
I am watching usage like a gas gauge now.
An agent loops. Every loop costs. If the default is to dump the whole payload back into context, a serious task is expensive even when you already did the cost engineering.
I am glad I have 4 systems on the same operating layer. When one meter is low I move the work and I get about the same consistency and results. That only works if you own the harness.
If you are buying AI work, ask what happens when one task eats a third of the week.