Your entire post is a huge red flag that tells me you don't know anything you are talking about just based on the first paragraph.

There is no qualitative change whatsoever. The training dataset now contains more agentic sessions and that alone was enough. Most model architectures that came out this year haven't meaningfully changed anything in terms of making models more expressive, if anything they took steps backwards in the name of inference optimization.

So it literally is better models in the form of better data and nothing else, otherwise it wouldn't have been so easy for competitors to clone leading capabilities.

Also LLM just means it is a large model that can output language and technically the LLM archetype was replaced by almost every new model being a VLM (vision language model) which is a meaningless nit pick.