GLM 5.2's reasoning effort dial can slash costs 20x

Zhipu's open-weight GLM 5.2 model offers frontier benchmarks at about a sixth of GPT-5.5's per-token price. However, actual task costs vary by over an order of magnitude depending on the reasoning effort setting, with the default being the worst. Properly tuned, GLM 5.2 can be both cheaper and more accurate than frontier models on easy and hard tasks.