Smaller OpenAI model excels at structured tool-calling and agentic tasks but is highly prompt-sensitive, raising questions about benchmark v