I think v3.5 is already a lot worse than earlier this year and v4 is basically the only option if you want a reliable answer.
This coincided with the update that made v4 the default option for new chats for Plus subscribers (of which I am one), drastically increased its speed, and removed the hourly message limits.
I hang around the OpenAI forum every day. The common sentiment this month is the opposite - ChatGPT v4 is worse, and v3.5 is actually better for many use cases. 3.5 handles roughly double the input tokens, so it's a lot better for programming and seems to be better at holding long conversations too.
There still are message limits too, it's just double.
so I am not crazy, 3.5 seems much dumber than before. It used to be able to solve coding problems and have a 75% working solution, now it spits out literal nonsense or doesnt even try.
This coincided with the update that made v4 the default option for new chats for Plus subscribers (of which I am one), drastically increased its speed, and removed the hourly message limits.