Benchmarks miss the thing that actually matters for agentic use: how does behavior change over a multi-day horizon? A model that scores well on one-shot coding tasks can still make terrible decisions when it has persistent state and resource constraints. That's where you see the real gaps between models.