Agreed, I reproduced the test and noticed that the second query (that used a index scan) had a weird estimate for the output. Image attached.

Actual number of rows = 70000 (scan so presumably reads first 70000 to get rows 50001 to 70000 as required)

Estimated I/O cost = 0.238681

Estimated Operator cost / Subtree cost = 0.0036486

Estimated Number of rows = 100!!!!!

So estimated operator cost = 100/70000 * 0.238681 = 0.0036484 (approx)

Don't know where the 100 estimated number of rows comes from.

Fitz