I remember testing Oj a few years back and finding that yajl where still faster for some cases, I believe it was very large arrays of thousands of objects.

I did some impromptu benchmarking now, and couldn't find a case where Oj performed worse than yajl. Impressive!