Whilst moving from filesystem v2 to v3 on Windows I found that the library uses CP_THREAD_ACP to convert between narrow and wide filenames. This is wrong, as I will attempt to explain. Background: Windows uses a number of locales and language settings as described on the following page: http://msdn.microsoft.com/en-us/library/dd319088(v=VS.85).aspx The relevant ones are Language for non-Unicode programs (System Locale), Standards and Formats (User Locale) and Thread Locale. The System Locale defines which code page is used by the system for narrow chars, and is set by the user system-wide. The User Locale is for formatting numbers and sorting and is set by the user system-wide. The Thread Locale is also for formatting numbers and sorting, is per-thread, and defaults to the User Locale but can be changed by the program; however it is recommended in a number of places not to touch it (although the exact reasons seem vague, apparently it can have surprising effects). CP_ACP gives the System Locale code page. CP_THREAD_ACP gives the Thread Locale code page which is rather a strange item considering that the Thread Locale isn't meant for code pages at all! "ANSI" (ie narrow char) API functions (including CreateFileA) use the System Locale code page (CP_ACP). fopen (VC++9.0) just passes the filename through to the Windows API and therefore also uses CP_ACP. fstream (VC++9.0) uses mbstowcs_s, which uses the global C locale (not the global C++ locale). Narrow char windows use CP_ACP. boost.filesystem should therefore either use CP_ACP (to be consistent with the Windows API, fopen and the GUI) or the global C++ locale (to be consistent with fstream). (Using the global C locale seems a strange idea, but since setting the global C++ locale automatically sets the global C locale as well, that seems a good compromise.) Using CP_ACP gives something that "just works", it's almost certainly what is needed. Using the global C++ locale is perhaps theoretically correct; it can be set (by the user) to match CP_ACP as follows: std::locale::global(std::locale(str(boost::format(".%||") % GetACP()).c_str(), LC_CTYPE)); which is likely the correct thing for most programs anyway, and also allows the user to choose something different should that be necessary. (Unfortunately std::locale("") doesn't do the right thing - it takes the user locale for everything, including code page, and as explained above the user locale has nothing to do with code page.) In many (perhaps most) cases the user locale and system locale will be the same, but not always - for example I have the user locale set to English UK (for dates) but the system locale set to Hebrew (I need to be able to use Hebrew in non-Unicode programs). Yechezkel Mett