Ask HN: What was the last task where only a frontier model could do it?

I have been seeing recurring claims that open models six months behind the frontier are sufficient for most work. I am asking for concrete examples from the last month where models like GLM, DeepSeek, Kimi, or Qwen failed while frontier models like Opus, Fable, or GPT succeeded. Please share specific tasks, the failures of cheaper models, and whether a slightly older frontier model would have been adequate in hindsight.