I’ve found Cursor to be hit or miss, but mostly miss in a big TS codebase. That is regardless of the underlying AI model. Claude Code, though, has been excellent so far. It seems the agent matters more than I’d realized.

Pretty much doesn’t matter in my experience.

If it’s a model you’ve heard of, it’s pretty good and pretty much the same as the other ones you’ve heard of.

Claude is still king for coding. Anything else I'm reaching for Gemini these days. The 1M context and the speed makes it way better for writing docs and explaining code.

Claude 3.7 is a little overeager. It'll solve problems you don't tell it to solve. Like I give it a screenshot of an app and ask it to do a button; it'll do the status bar as well. It's somewhat like a jerk genie and I feel like I'm writing a contract rather than working with an assistant. Claude 3.5 doesn't have this problem.

Gemini 2.5 Pro will one shot complex instructions and even read between the lines. The drawback is that it can often overthink as well. One example was I want to deselect a dropdown if I recreate it (e.g. select state when changing countries). The ideal solution is to just check this on the view layer. Gemini 2.5 pro is terrible at this - it'll go down into lower layers and overengineer, when the solution should just to set as null if the new dropdown doesn't contain the same text.

GPT 4.1 has been quite good lately, but it's not quite as agentic and it will ask you a lot of questions and request confirmation to begin. It is perfect for workflows where you aren't certain yourself.

GPT is also smartest from the visual angle, able to read color variations well and designs them well. Claude is the worst visually.