cost of e2e task resolution should be cheaper, even if single inference cost is higher, you need fewer loops to solve a problem now

Sure, but for simple tasks that require a large context window, aka the typical usecase for 2.0 flash, it's still significantly more expensive.