Best model?

[TheLazyCat00]

What do you guys think is the best model currently? Theres so many options:

o3 vs o4 vs gpt 4.1 vs gemini 2.5 pro etc.

[igorgue]

For CodeCompanion to me the only one that does use the tools correctly and most consistently is good olde gpt-4o. I just tried gpt-4.1 and it never triggers the editor tool, same goes to Gemini's if you ignore the tools usage though, I think Gemini 2.5 pro is the best, followed by DeepSeek R1 Reasoner, in my experience if you give DeepSeek a lot of context and makes it reason it works better.

[olimorris]

#1141 is around the corner and I have list of adapters and models that are fully working with tools. They may need some extra prompting to get the job done but they're functioning well.

Any Anthropic model has been super reliable for me in most tasks especially agent use.

[TheLazyCat00]

I've been using 2.5 pro (from copilot) and honestly it works pretty fucking good. It also makes very nice code and seems to have a good understanding. But i cant say that its the best cuz i didnt use the other models as much.

[jinzhongjia]

copilot gemini 2.5 pro and claud 3.7 are very great!

They can work very well.

But there is a problem, when I user mcp file tool(handle files), the response xml always parse failed

[ravitemer]

Yeah! Neovim MCP sever's file tools often run into xml parsing issues with the xml version of tools. codecompanion is being migrated to function calling very soon. It should be fixed with that.

[pratyushmittal]

gpt-4.1 and o3 are very special.

• This post by Simon Willison is highlights the search capabilities and accuracy of o3. o3 can do a web-search when it cannot find the answer otherwise. This helps in rust and other codebases where current models are otherwise not very good.
• gpt-4.1 is specifically trained for function calls. It has outperformed all other models in picking right function tools. It can call function tools recursively. This makes it special. Plus it is very very intelligent.

I would go for o3 if we can use search with it.