GLM 5.2 beats Claude Opus 4.7 in benchmarks: code generation and reasoning tested