It's likely that you'll get reduced performance from this, as blowing up the token count is part of getting a marginal amount of increased performance out of the models.

Funny though, as is most discoveries related to emergent LLM properties