I think it's both. It's easy to run up a massive bill with AI without much to show for it, which is why token-maxxing has now been replaced by cost-awareness and AI budgeting such as Uber's "max 10% of salary per employee".

Cost isn't just token price though - it's (number-of-tokens-used x price-per-token), and there are large differences in token efficiency between different models and harnesses. Increasingly we're seeing benchmark sites focusing on "cost per completed task" as a cost metric, and it's not always the cheapest tokens that win.

I agree that ultimately AI/coding cost is just part of the picture - at the end of the day it's about software development cost, which for time being involves humans.