Not if you have a 5-6 second F2L like Feliks, and when I do use an OLL algorithm, it's only about a second or two faster at most compared to 2-look OLL.
How did learning all 57 OLLs drop your time by 3 seconds? [F : [R, U]] and [F : [R, U]'] barely takes one second to execute. Plus, there's a factor of recognizing the OLL cases, which is a bit more difficult than recognizing OCLL.
Maybe more like 2 seconds. The difference in the recognition time for all the different OLLs is about the same for me. Recognition + possible cube rotation + execution + possible no edges case which results in a having to perform the T-orientation twice = 1-3 seconds of extra time.