Are Execution-First Models Getting Underrated for Agent Workflows?

A lot of model discussion still gravitates toward benchmark screenshots, clever chat demos, or long reasoning traces that look impressive at first glance. Those things are easy to share and easy to evaluate in isolation. But once a model is actually embedded inside a product or an agent workflow, I am not convinced those are the most useful ways to think about performance anymore. The question I keep coming back to is much simpler: how much useful work does the model actually get done per token,