This. They've all reached the point where they can be decent coding assistants/rubber ducks. They can all also do a good job at general stuff like helping me write my emails, answer basic queries etc.

The only "value" the cutting edge models provide is if you're looking to hands off and trust the models to complete full tasks for you or implement entire features. In that sense some models are better then others. Some will give you a working solution on the first try. Others might take a few tries. The problem is that none of them are to the point where you can actually trust their outputs. One model being 10% or even 2x more trust worthy with its outputs isn't meaningful because we need orders of magnitude level improvements before we can begin trusting any of these models.

And anyone that thinks any of these models are reaching that point right now is likely ignorant of whatever subject they're having the LLM generate code for. I haven't gone a single coding session with any of the top models without spotting subtle but serious issues in their outputs. Stuff that if I caught once or twice in a code review, I wouldn't think twice, but if it was daily? I'd be looking at replacing that developer.