Panopticlick claims that my user agent (Firefox 4.0b7 on Windows XP) is about 1/15,000. That seems unlikely.
It claims that my http_accept (text/html, STAR/STAR ISO-8859-1,utf-8;q=0.7,STAR;q=0.7 identity en-gb,en;q=0.5 -- except that I've replaced asterisks with "STAR" to avoid HN formatting confusion) is about 1/19,000. That seems even more unlikely.
I suspect these have the same cause: they have this big database of browser information, gathered over time, and recent browsers are underrepresented. So anyone running a Firefox 4 beta, or a recent Chrome build, will show up as being very unusual in the database, but that's misleading because some of the older entries in the database will represent browsers that are no longer in the pool, whose users are now using something more recent.
On the other hand, I don't find it so hard to believe that my browser plugins give 20 bits or so of information. (Though some of the same bias will affect this figure.)
In case it isn't obvious, by the way, their figures are simply the fraction of records in their database that have the same user agent / browser plugins / system fonts as yours. The browser plugins figure is 1/1,277,632 (or, for me a little later, 1/1,277,946) because that's how many times the Panopticlick site has been used.
Hmm, I wonder how they defend against having the same user run their test twice in a row. (Answer, having tried it: they don't. Probably a good thing since the obvious non-privacy-compromising way could skew their figures a bit.)