Are you implying they can’t do both?

There is a theory that the verbosity and comments help getting better results with the current benchmarks. So the models are theoretically getting better but in practice they are getting worse.

Theyre training the system to minimize compute,so most likely theyre dynamically downgrading quants in the first few turns hoping to find the cheapest model to run. The side effect may be excessive token gen

You're joking, right? That's an incredibly outmoded view of capitalist "competition."