“The Chinese model GLM 5.2 outperformed GPT-5.5 on the frontier SWE coding benchmark and trails Claude Opus 4.8 by less than one percentage point.”