My suspicion is that some training environments had a length penalty based on word count rather than character or token count.