There are two likely reasons: efficiency and "alignment".
To make models more efficient, you have to reduce computation burden. This can be accomplished via quantization and pruning, both of which will reduce computation at the expense of perplexity.
Alignment (censoring by another name) also reduces perplexity as we attempt to coerce a model from its ground state into some more socially acceptable version.
Google has an ingenious approach whereby they debuted with a bad model, and are secretly improving it. OpenAI did the opposite, they released a model that blew everyone's minds and are now making it worse.