Why is claude 3.5 sonnet superior for composer?

I see a lot of praise for o3-mini models, but I can’t seem to get them to work properly for the agent mode.

When I ask sonnet to do something, it looks for context to try to understand the relevant files. O3-mini just works what context is in the file I’m working on, giving far worse results.

For me, Agent mode has been a gamechanger, and it seems like sonnet is the only way to go for now.

1 Like

Yeah, I suppose it largely depends on what you’re working on and what you expect to get out of the agent.

I’m with you, I find Claude Sonnet to be such a much more reliable and consistent assistant. it will look for context, it considers everything you say and tries to address all your points.

o3mini likes to give info-dumps and then often has to be begged multiple times to actually write code. It doesn’t seem to be as helpful as Claude.

I’ve had it once where it scanned through a dozen files one by one “getting context” and I’m thinking OK here we go it’s going to do something amazing… Then in the end, it literally changed 2 words on 1 line in 1 file. Needless to say, the change did nothing to address the issue I asked of it. :rofl:

1 Like

Its clear that the prompting has been optimized for non-test time compute models. I find o3-mini to be amazing in chat, but when it comes to composer, we often get erroneous references to files, code just showing up as text, etc. However when it comes to o3, the longer context window, and output tokens really is amazing. I hope Cursor works with OpenAI to optimize it for composer because from my experience coding this weekend, it is a more powerful model.

Not sure OpenAI will collaborate very much, OpenAI seems to be finalizing their own coding agent probably leveraging o3-mini