I agree with you, but then I am retired so my opinion is not that relevant.

I have a sweet spot for using just Emacs, no other IDEs except very occasional use of AntiGravity. For a particular fun project or researching how to use generative models in applications, I like to start low: see if a small local model with appropriate tool use or agentic library will get the job done, if not move up to using something cheap like gemini-3-flash, and only if none of these approaches work, then use an expensive model.

I was advising a friend’s company last month on their application that effectively uses LLM models, but I was blown away by their zeal to spend lots of money on tokens.