It is certainly true that two plans will be built ?

Are the plans built using exactly the same input parameters ?

If not that could explain the difference, plan a gets built using the value of 'X' as a predicate where plan b uses 'Y'.

That can make a big difference.

Im not aware that artihabort on itself will make the optimizer make a different choice in producing the plan.

Are you able to provide the plans for both queries ?

If you chop and change the arithabort setting in SSMS does the query 'magically' go faster and slower ?

What if you clear the statement cache before each execution ?