Personally I found gpt-5 to be a bit better than sonnet-4. At least in cursor. Claude is still more reliable and competent at tool calling, but I found gpt-5 to be better in token efficiency and a lot better at instruction following.

This seems to change every 6 months? Just my impression.