Claude is decent for sure, but if you are using these models for 'smarts', that is a whole separate problem. I also think honestly people are sleeping on Mistral's medium 3 and devstral medium. I know it isn't 'smart' either (none of them are), but for mundane tasks need valid code output, it is extremely good for the price.
As a sounding board for things you are already well familiar with, I agree, and have experienced the same, and that can be useful. It's also a much better experience than say using Google to do the same, or just a rubber ducky.
The NLP these models can do is definitely impressive, but they aren't 'thinking'. I find myself easily falling into the habit of filtering a lot of what the model returns and picking out the good parts which is useful and relatively easy for subjects I know well. But for a topic that I am not as familiar with, that filtering (identifying and dismissing) I do is much less finessed, and a lot of care needs to be taken to not just accept what is being presented. You can still interrogate each idea presented by the LLM to ensure you aren't being led astray, and that is still useful for discovering things, like traditional search, but once you mix agents into this, things can go off the rails far too quickly than I am comfortable with.
