Generally, whenever I do see frontier AI model's evaluation metrics on say Artificial Analysis. It gives out a benchmark on a set of questions (or tasks). But it never quite actually shows what tasks it fails at. I'm more curious on benchmarks that focus on SWE.