GLM Beats DeepSeek-V4 With 5x Fewer Tokens On Voxel Tasks

I hope they don't stop here GLM 5.2 is still staggeringly better

@teortaxesTex The user? обычно так glm начинает размышления, а deepseek v4 начинает We need.

@teortaxesTex tested it too, feels very unpolished, better tool calling but metr would be really bad and self iteration feels impossible