The crosswords are tough for a stock LLM to crack without invoking tool calls since they can hardly spell due to their tokenized inputs. But with tool calls, they’re getting pretty good at it.