The gaps between models is always shifting so this will change very soon, but right now GPT-5.2 Pro clearly stands out for hard problems, Gemini 3 is the best model for Deep Research (especially in NotebookLM) and Claude Code/Desktop/Cowork powered by Opus 4.5 is best local agent

The good thing is no matter which of the Big Three you use, they are very much neck-and-neck, and if you give it a couple months, they inevitably each release versions of the other's products (though NotebookLM has been mostly unchallenged so far).