What's Next for LLMs?

I've been using DeepSeek v4 flash for guided agent workflows (in-IDE, prompt/review diffs) and have found no noticeable benefit to using Haiku, Opus, Sonnet in this workflow.

For around a dollar a day, I am able to more than double my own productivity. What are the frontier models for/what are they trying to solve?

Are they designed to one shot applications? Is the expectation that we want to be able to "yolo" prompt LLMs and have them complete work unsupervised?