Is there really a difference in coding performance between Claude 3.5 and GPT-4o?

Facts — Claude is really good at hanging onto long files.

I also felt GPT was better at clean algorithms lol I thought it was just me

But isn't that old news? Since the recent models came out, it feels a bit different in practice. Just looking at benchmarks, the rankings keep flipping, so I think the bigger factor is how well you organize your prompts and context, rather than the model itself.

In practice, I just use both. I throw the big-picture stuff—design and refactoring—at Claude, and ask GPT for the small things, like writing a specific function or a regex. If you lock yourself into just one as your main, sooner or later you'll be fully exposed to that model's weaknesses, so alternating between them ends up being easier.

If you're trying to hook up code review automation, you need to think about tone too. Claude's review comments sound human, so teammates push back less—but when it comes to inventing bugs that aren't there, both are exactly the same lol

I ran a similar test, and the conclusion was "pipeline over model." When I included not just the diff but also the related files and test code together as context, the difference between the two models shrank dramatically. And for verifying the results, running another model again for cross-checking was the most practical approach.

Got a source? Leaving aside benchmark numbers, real-world feel is all subjective lol

I agree, but it probably depends on what counts as an 'algorithm problem.' For Baekjoon-style problems, GPT makes fewer syntax errors, while for refactoring actual repos, Claude is much better.

I paid for both and tried switching between them month by month, and now Claude is my main. That’s because with long files, it’s definitely less likely to suddenly forget earlier context midway and start spouting nonsense. GPT is snappy, but sometimes it’s confidently wrong.