I just created a post on the same subject a few hours ago. I have found much the same as you, that the thinking models are not actually better than the non-thinking, at least for my regular day to day tasks. All three of your key points, are true for me as well.
I’ve been using gpt-5-fast today. Its a thinking model, and its been a lot slower. I’m trying to see if it is a better coding model…yesterday was not a good day with the gpt-5 model. The reduction in my overall speed with the thinking models is definitely frustrating. The gpt-5-fast model is faster than sonnet, and when the LLM is actually doing real work, its speed is nice. But the periodic thinking processes really hamper things, and I’m generally slower today, than when I use claude-4-sonnet (non-thinking.)
One thing I DO like about the GPT-5 models. They are not egregiously over-agreeable, over-positive, overcompensating with kindness, even in their thought processes. I don’t mind kindness, but when you are telling a model it screwed up to a mind-boggling degree and it needs to fix its **** and it says “You are absolutely right! Of course! I need to be less of an idiot! HAHAHA!” it gets a little annoying. 