Agreed, these types of posts often feel like they're missing the forest for the trees. Sure, migrate away from Claude and maybe that will provide some runway, but all of these companies are built on the same economic fundamentals that do not scale.

We are currently in the "$7/mo Netflix with all the good movies" era of AI that will leave and never return.

Or LLMs become a commodity if open models are ever good enough for > 95% of use cases. The product if fundamentally different than tv-shows or movies since its so interchangeable.

Then the cost would end up being deprecation + electric + some low operating margin (i.e. what non SOTA models cost on OpenRouter)

That's why I'm switching to open weight models. I'm running models locally, self-hosting and using third party for various things. Like it says in the article each model has its strengths and none of the proprietary harnesses allow you to orchestrate different models depending on their strengths. It's cheaper, more flexible, private, more resistant to rug pulls, and brings back some fun to building for me vs just auto-accept, yell BAD CLAUDE when it breaks, repeat ad nauseam.

Its more like Netflix being the only service with movies available at 8k, some other options having movies at 4k and then a free option where they're only at 1080p.