Is it safe to generally assume that toString() has a low cost?

The best way to find out is to profile your code. However, rather than worry that a particular function has a high overhead, it's (usually) better to worry about the correctness of your application and then do performance profiling on it (but be wary that real-world use and your test setup may differ radically). As it turns out, programmers generally guess wrong about what's really slow in their application and they often spend a lot of time optimizing things that don't need optimizing (eliminating that triple nested loop which only consumes .01% of your application's time is probably a waste).

Fortunately, there are plenty of open source profilers for Java.

My thought is:

Yes on standard-library objects

No on non-standard objects unless you have the source code in front of you and can check it.

There's an easy answer to this one, which I first heard in a discussion about reflection: "if you have to ask, you can't afford it."

Basically, if you need ToString() of large objects in the day-to-day operation of your program, then your program is crazy. Even if you need to ToString() an integer for anything time critical, your program is crazy, because it's obviously using a string where an integer would do.

ToString() for log messages is automatically okay, because logging is already expensive. If your program is too slow, turn down the log level! It doesn't really matter how slow it is to actually generate the debug messages, as long as you can choose to not generate them. (Note: your logging infrastructure should call ToString() itself, and only when the log message is supposed to be printed. Don't ToString() it by hand on the way into the log infrastructure, or you'll pay the price even if the log level is low and you won't be printing it after all! See http://www.colijn.ca/~caffeine/?m=200708#16 for more explanation of this.)