What you need for real engineering and scaling is cheap and just enough smarts. I've found GPT 6 Luna to be cheap enough that I can pretend the cost doesn't exist and it's smart enough that I can get real work done. This let's me scale it in ways that I can't scale frontier models. And funny enough, the LLM I have running on my 5070 Ti is just about comparable to what I get from Luna. Unless frontier intelligence gets waaaaaay cheaper I basically don't care about them.