The costs are estimates, they are always estimates. The only thing actual about an actual plan are the actual executions and actual row counts and a couple of other actual counters.

Why would it? The costs are there for the optimiser. They're how the optimiser tells which plan appears to be the best of the ones it found. It's a cost-based optimiser. Once optimisation completes and the query's execution starts, those costs are no longer needed for anything. So why would SQL go to all the expense and effort of calculating run-time costs when the costs are only needed before execution starts?