A lot of people are using Claude Code which many consider to be a noticeably better for coding than the other models.

I think also they tend to be generating non-C++ code where there are more guardrails and less footguns for LLMs to run into. Eg they're generating Javascript or Python or Rust where type systems and garbage collection eliminates entire classes of mistakes that LLMs can run into. I know you said you don't use it for Python because you know the language but even experienced Python devs still see value in LLM-generating Python code.