Yeah, I think when they made the bet it genuinely made sense. But in coding workflows, once models got cheaper, people did not spend less. They just started packing way more LLM calls into a single turn to handle complex agentic coding steps. That is probably where the math started to break down.