It's not too much work, although things like unquoted tokens do introduce some ambiguity (a technical term for when two productions have overlapping start symbols) in the grammar. This is usually an academic concern though, as ambiguity can be resolved through backtracking, and as long as you don't require too much of it, it's fine.
I've implemented a hand-written recursive descent parser for something that is very similar to JSON5, and it wasn't difficult overall.
But all of my services speak JSON, so if I can do a tiny bit more work to speak the same language, that's an advantage that JSON5 loses over JSON. I don't think JSON5 is enough more human-friendly to make up for that rather large disadvantage.