ORCA benchmark trips up ChatGPT-5, Gemini 2.5 Flash, Claude Sonnet 4.5, Grok 4, and DeepSeek V3.2 In the world of George Orwell's 1984, two and two make five. And large language models are not much better at math....