Just to point one thing out even more:
If you look at the results, === performs about ten times
faster than == when types are different.
That's not as much as, for example, the complexity of different sorting algorithms where you go from O(nČ) to O(n * log n), but a factor of ten is a huge improvement. And it's as cheap as one additional "=".
Maybe the main argument against my test is now going to be that "5"==5 is true and "5"===5 isn't, so they produce different results which may not be wanted. And that means you'd need to convert before using ===. Which will slow it down again.
However, the following test shows that even if you only compare once, conversion + === is still faster than ==. Using more than one comparison, the initial conversion will outperform == even more: http://jsperf.com/versus-conversion