Isn't it a lot more than 3%? I don't think I've heard anyone say url-safe-base64 and actually mean urlencode(base64(x))

… yeah. I assume they're getting that from doing 3/64, but for uniform bytes, you're rolling that 3/64 chance every base64-output-character. (And bytes are hardly uniform, either … TFA's example input of JSON is going to skew towards that format's character set.)