I'd argue that even GPT-5 and Sonnet 4 at their highest reasoning budgets are not enough for "real coding work" because you still have to think about how an LLM should do it instead of what, and put it into a prompt. some harnesses, such as JetBrains Junie or Gemini CLI, make a good job of letting me drift into declarative prompts, but that's still not enough.

totally agree - they all need a human in the loop at this point. I'm constantly stopping gpt-5/sonnet 4 and steering. Unfortunately with grok it completely misses the plot constantly