If I used Babel or any other transpiler for critical production code, I would only enable a whitelist of features that I thoroughly understand: for example, how a fat arrow is translated into a regular function. This vetting would be time-consuming, but that's a price I'd be willing to pay because I enjoy writing ES6 much more than ES5. That's highly subjective though, so I completely understand how some other people would rather just stick to ES5 and get stuff done with the lowest possible risk of weird things happening.
I have an affinity for Babel unlike any other JS project (except React). They team works really hard to keep it interesting.
For production code I'd still stick to a whitelist of features that have been released for at least a month though. Babel is aggressively pursuing ES6 coverage and new features might not have all edge cases covered when they're released. Give them one month though and with their responsiveness it's 90% likely it's going to be fixed.
Besides all that everything ever created has bugs; why risk introducing bugs from a JavaScript engine and from a transpiler when you can risk a single one?
As an aside my first and last time using CoffeeScript I ran into an issue that, since this was about two years ago, has already been fixed but at the time it cost me hours of debugging time. It really soured my experience with any transpiler.