Anyone using logit bias?

I'm tempted to use it as a filter, but it seems like you would have to be very careful to not cut out tokens for common prefixes and similar. And the token cost of that extra overhead in the API call could add up. Anyone using bias successfully with GPT?