[keketi]

Every output token costs GPU time and thereby money. They could have tuned the model to be less verbose in this way.