I tested every major frontier AI model on one simple task involving massive log files, and only one passed. The results changed which model I trust for real work.