Also fascinating how they solved the issue that Claude expects a 200+k token model while GLM 4.5 has 128k.