aichannel

Question for devs: is Claude still the best model for actual coding?

If so, by how much?

I've tried r1, o1, and o3 and the answer is still clearly yes

Claude is no longer my go to for big architectural discussions, but it's the best at understanding a codebase and delivering a clear solve with minimal refactoring. Oh, and it's still blazing fast in comparison.

So no for high level, yes for code

co-founder @survey x /ponder // benadamsky.com

o3minihigh will be a great replacement very soon

prev founded rocketr.net, swish.id now exploring aiXcrypto identity ai. expired sprint car driver & energy data engineer. idol worship on shoni_eth.X

We shall see, been surprisingly hard to beat sonnet even with the latest models though

Expecting a big breakthrough in the next couple of months, but so far nothing has been worth making the switch even though sonnet is a mid level programmer at best

ah i don’t agree at all, imo sonnet is so bad now i sometimes just use 4o instead. total lack of instruction following

on numerous occasions in a big codebase (elizaos, 800 line files) it literally didn’t do any code and just made a comment note changing a word or something ridiculous. 

are you using cursor agent?

If you want really good boilerplate based on a lengthy spec than there are better models than sonnet

For me, I'm typically looking to prompt for specific high level output and claude has been better at that than any other models I've tried

Tried cursor agent but have a better workflow w/ chat tbh

As in when I'm building I'm usually looking for quick output on a small prompt, not instructional prompts w/ a lot of automation which seems like what cursor agent is geared towards

For what you're describing I think sonnet is a much less capable model, I wouldn't use it for that either

The biggest learnings for me working w/ AI is figuring out which models are best at the specific things you're looking for, many of them have a specific edge

workflow of not switching to o1 or wdym? agent really just automates cli stuff i have to copy and paste otherwise

re boilerplate: i am making core edits to major files and it just is not capable at this level. it was fine on smaller projects— idk i just have no interest in it anymore it takes more time to use than is worthwhile