I was using DeepSeek Coder and Claude for verilog quite successfully back in June - saved me a lot of time creating testbenches. Probably better now. I need to try the latest Gemini 1.5 pro to see how it does. I used DeepSeek Coder to create a testbench for some legacy verilog code that was poorly documented - it did quite well, better than I was expecting. It even collected repeated code into tasks which kind of surprised me.
I hadn't even considered trying to use ChatGPT with Verilog. I strongly suspect it will get in the way. My design is pretty darned small, just a grid of uniform cells, with not much else in the way of I/O, etc.