I've mostly switched to using Claude these days, with MCPs for websearch and fetching specific remote or local files. It answers questions generally very accurately (from the source documents it identifies) and includes citations.

I've found that people that haven't really tried the latest models, and just rely on whatever knowledge is in the model training are really missing out on the potential power. GPT4o+ and equivalent models really changed the game. And using tools to do a search, or pull in your code, or run a db query or whatever enables them to either synthesize information or generate context relevant material. Not perfect for everything, but much better than a year ago, or what people are doing with the free systems.