Have you noticed a difference in code quality between Claude and GPT?

Agreed—Claude definitely breaks things less when you iterate on edits. GPT starts calling functions that don’t exist by around the third pass lol

Having used both in real work for about half a year, they definitely have a different feel. GPT is really good at cranking out an initial design, while Claude is strong at gently refining and refactoring it. So I draft with GPT and clean things up with Claude. But this can flip depending on the model version, so I think 'always Claude' is a bit risky.

Isn't that outdated? With recent models, GPT also handles long contexts well. What you're feeling might just be a difference in prompts.

Any source? If there's a benchmark, I'd like to see it.

In practice, I use both. If I only use one, I get stuck in that one’s habits. I use Claude for refactoring and GPT for generating test code—I split them up like that.

I mean, isn't making up nonexistent functions just a matter of how you prompt it? lol Just adding one line like 'Don't create functions that don't exist in the existing code' cuts it down dramatically.

A function that doesn't exist on the third revision lol. This is so relatable—it's funny, but it hurts.

Let me drop one tip: when you have it refactor, don't let it spit out the entire file again—try asking for only the changed parts as a diff. That way it does less “making stuff up” (?). And the biggest thing is splitting commits into small chunks and having it make only one change at a time. This workflow makes a much bigger difference in practice than model differences do.

Honestly, concluding that “Claude is better” based only on the author’s personal experience seems a bit premature. A sample of one side project is too small to generalize from. Many other people have had the opposite experience as well.