There is another possible approach, but I haven't decided yet whether or not I like it.  The email package string parser could (and may for other reasons) become smart enough to convert unicode into the charset declared in the MIME part when it is parsing the string version of a message.  In that case, smtplib could use the string parser to parse the DATA payload, and if it parses successfully it can then use the same code path as I'm proposing for send_message to generate the right output depending on whether or not the smtputf8 capability is present.  That would place a new constraint on what was acceptable as a DATA payload, though, so I'm not at all sure it is a good idea.