My team has been using gpt-4o as LLM judge for a specific usecase; it just bad on evaluation set after iteration on prompts. But switched to claude 3.5(latest). Oh boy. From CoT and the performance it just gets it. Performance is consistent at 95-98% on eval set.