This very much reminds me how the earliest users get a bit screwed due to high costs and low quality. Maybe these users are caused enthusiasts, so they don't care. The next set of users get higher quality and lower costs. These are your big winners in the timeline of a technology.

The middle users get benefits, but aren't treated quite like the previous set.

Then you have your late adopters, these are the ones that are lightly abused, but they get a mature product, so it isn't that bad.

Finally you have your last users. These are milked for every drop. They have irrational loyalty or are locked in.

I imagine AI will follow this trend and we are entering past those middle users as we speak. Seeing how little difference there was between GPT 3.5 and 4, and how computationally expensive 4.5 was, I think we've hit the end. Now its just how many prompts do you want to run for COT/Thinking?

GPT 5.2 is cheap and they are proposing ads. They see the end is near and need to capture profit before local models take over.