MagicDraw:
Pros:
Supports the OMG specifications flawlessly (at least, it seems that way).
Easy to use.
A lot of capabilities and very innovative in what they do (their new product where someone can access the model online and write comments on it is genius; Workbench is also kind of brilliant).
Can connect with lots of other tools (e.g. DOORS, MATLAB, etc.)
Cons:
No Magic kind of nickel-and-dimes you to death. You get MagicDraw, but it only comes with UML; gotta pay extra for UPDM and/or SysML (and whatever other features you may want).
To effectively use the report generating features, you have the learn Velocity Template Language and the specific library for it for MagicDraw.
I hate their implementation of stereotypes (at least, as I would want to use them). Ideally, I would make a profile and make some metamodels and, at best, use those stereotypes on any diagram I want or, at worst, have to set which diagrams I want to use those stereotypes on. In MagicDraw, I have to make new diagram types to use stereotypes, which I find kind of goofy. What if I just want to use my stereotype on a BDD? Why do I have to make "BDD w/ Stereotype" as a diagram type?
Sparx EA:
Pros:
Probably the best user interface of them all. Very easy to draw diagrams.
Everything comes in the package.
Cheap.
I really like how they handle stereotypes. It takes some work, but it works well once you get the hang of it.
Neutral:
Reporting is kind of weird. There is a GUI for making report templates, but it is severely limited. So you turn to package fragments and selectors, and then you find that those are a bit limited too. So then you turn to SQL statements, but now those are limited. So you finally throw your hands up and start writing scripts for your reports, at which point having the GUI for it seems completely pointless. The bizarre thing is that the changes that could make it so much better are so minor and (IMO) obvious that it's baffling why they aren't in there. There was a good idea there at some point, but it just never quite got there.
Cons:
Lets you do whatever the hell you want so it flies off of the rails as far as following the specs (granted, this likely feeds my Pro regarding stereotypes).
Kind of an extension of the above, but if you're a SysML purist trying to stick to the standard usually takes you through some crazy unintuitive rabbit chases (if it works at all).
Not many features when compared to other tools. EA is very bare bones in my experience.
For some reason EA maintains graveyards of relationships it seems. For instance, in our EA model one day we deleted a bunch of stuff, and somehow the model size grew by several MB (maybe up to 30, I can't quite recall). It seems (though we can't really verify this) that when you delete an element that's in a relationship with another element, the relationship hangs around and somehow takes up more space. What led us to this conclusion was when you transferred a huge EA project to MagicDraw and got 50,000+ errors; most of these were "dead" relationships).
In general, as a SysML purist, I generally find working in EA a bit frustrating. It just does so much weird off-the-wall stuff. In fact, overall I find working with EA frustrating. You'll have a clear idea of how to do something, and then it just completely does not work the way you think it should work. This was especially true when I was learning the document template functionality.
Rational Rhapsody:
Solid user interface.
Seems to follow the specifications fairly well. I haven't really encountered anything that sent up red flags, but I've used it considerably less than MagicDraw and EA.
Absolutely fantastic document template GUI. They also have a $3000 add-on that supposedly does even better with generating reports, but what comes out of the box is just killer, and I wish it was the standard for all of the tools.
Cons:
I think they're generally behind on updating to the latest specs. Could be wrong on that. Overall, I haven't used Rhapsody enough to find many cons.
Artisan:
I've only seen a presentation on this. It looked great. The way they handled stereotypes looked great. The reporting looked pretty decent. Matthew Hause is involved so there's no doubt in my mind that they follow the standard. The idea of Variant Modeling is absolutely fantastic as well. Unfortunately, I was never able to get a demo of the tool (hands on demo where I could test the tool for a couple of weeks) as the sales guy I was talking to quit responding to my emails/phone calls (admittedly, the folks making the actual decisions on my side were being kind of wishy-washy, as they'd say they wanted a rep to come down one minute and then change their minds the next, and I was just stuck as the messenger). Also, the tool (with the variant modeling) will cost you big money (somewhere around $40K I think) whereas you probably won't crack that for a small team for the other tools (maybe not even half of that; depends on how many plug-ins you buy from No Magic though). Granted, I could never figure out if that $40K bought enough licenses for many users or was just one user, I would assume the former, at which point it's still highly expensive but not terrible. Still, I'd love to try it out one day as it looked really good.