It would be reasonable to expect that extracting and parsing operations from an applyOps would be somewhat similar in cost to parsing the inner oplog entries if they weren't part of an applyops. Unfortunately this is not the case:

BM_NonApplyOpsParse/64/100                1044574 ns      1044562 ns          671
BM_NonApplyOpsParse/500/100               8120293 ns      8120136 ns           86
BM_ApplyOpsExtractOperationsTo/64/100     3752842 ns      3752796 ns          186
BM_ApplyOpsExtractOperationsTo/500/100   28395271 ns     28394950 ns           25

The parameters are the number of operations and their approximate size. ExtractOperationsTo is MUCH slower, and this may be the cause for some slowdowns in secondary application seen when batching inserts.