Additionally, on the context front -> all the labs are aware that for many tasks you can get 10x+ increases in output quality by feeding better context.

See https://arxiv.org/abs/2604.04364.

This won't really show up in benchmarks, but it will impact real world usage on the most common use cases.

I'm doing a study right now on the impacts of better context for small models to fix bugs.

A very dumb algorithm can make small models perform at 10x+ model sizes. I'll be surprised if it can't get to 20x+