They used to be remarkably good. Shortly after PG wrote his A Plan for Spam essay, I set up a customized system that used my own spam e-mails as training data. I can't remember the tool I used, but it involved a script on top of maildir which is sure to be far more kludgy than what you could do now. The technique is powerful enough I suspect it still works remarkably well today because it's tuned to your own corpus of spam messages. Nonetheless, for the last decade I've been too lazy to do this and just forward everything to gmail, though gmail does give me a few false positives every couple years.