We Ran Grok 4.5 Against Sonnet 5 and GLM 5.2

27 sandboxed attempts on real code-fix tasks. Same top score, very different bills