>token use is higher than k3 and far higher than proprietary models

GLM sets effort to max by default historically.