We just added support for the new Claude 4 models: Sonnet and Opus. With this launch, we’re offering them at a 50% discount for the next couple of days. We’ll make sure to announce pricing changes beforehand.
Sonnet : 0.5 requests for regular 0.75 for thinking
There are weird loops happening with Sonnet and Opus 4 where after making a change it thinks the formatting got messed up and keeps trying to fix it. This isn’t happening with 3.7 or any other model. I assume this is cursor-related and not the models just being weird? Do they apply/see changes differently (search_replace)?
Decided to try it on my C# project and Sonnet often makes formatting errors. If I understood correctly what I saw in his terminal, he makes unnecessary tabs and sometimes skips line creation or incorrectly positions edits.
Some of the problems he was able to fix with dotnet format, and the ones he couldn’t fix, he fixed with a complete rewrite of the file ¯\_(ツ)_/¯
Far as I understand it’s due to Claude 4 and Cursor’s Tab/whatever other model unable to work together cohesively at the moment. As it’s the Cursor’s own model that applies code changes, far as I’m aware. Not Claude.
Been using it all day – btwn claude desktop, v0, gibson, cursor .50…
Although cursor bungled my repo that v0 started for the ux, so I had to revert to an earlier version, but went too early - and v0 lost several versions, and I kind of messed up my deployment - so will have to redo:
Actually I’m fairly certain it is because Sonnet 4 and Opus 4 don’t use the same method of applying edits. Other models use the apply method where it summons another smaller custom model to apply the edits proposed by the model in the way you describe, but I think Sonnet/Opus 4 are directly using a string replacement tool.
I’m pretty sure this is also related to another issue that is specific to Sonnet/Opus 4 where the “restore checkpoint” button doesn’t work if the files that were edited were not manually included in context.
I’m finding that the new Sonnet struggles with Cursor’s tool usage in agent mode (Python). It will also keep making diverse formatting mistakes. This is for editing files though, it seems to do well enough at generating code from scratch.
It’s unfortunate and concerning that Cursor has made this article title unsearchable on their forum — this kind of censorship undermines open discussion and user trust.
