Z.ai's post-trained GLM-5.2 base scores near-frontier coding results at lower cost per task, but is the token-hungriest model in its class.