Its great. I use both. I do coding at all kids of levels languages and environments and when I want boiler plate with out thinking about it, its mostly right.

It won't solve obscure problems, but it's really good for generating lots of beginner-level code or answering beginner-level questions.

Significantly better code gen, but not to the point where it can make an application on its own. I tried using it for an embedded esp32 based project for fun, and while it could create mqtt support, the code for setting up WiFi / a small web backend / some HTML for a front end... It struggled with the application logic. Either way, it got about 70% of the way there

In my experience its better but not a ton better.

GPT 4 can reason better with complex problems than 3.5

Personally I use perplexity.ai with gpt4 only because the answers it generates are preceded by a web search it does from what you ask it. I found it to be far more accurate just because of this.