The default Balanced profile now places each run on the cheapest offering that fits and has a machine free. If
it cannot be used, the next cheapest takes the run, and so on. Health, startup time and fit only decide between
offerings within 2 % of the cheapest rate.
The estimate before launch quotes the rate of that offering, and the run is charged that rate, frozen as before.
Each offering’s from rate and host shape, and each card on the GPUs view, follow the cheapest machine free now.
When it sells out they move to the next cheapest, and back when it returns.
To rank on the expected cost of the whole run, startup and lost work included, set placement.profile: Cost.