Been using Cursor for work and Claude code for pet projects. IMO Cursor is better just as Carpathy said - It’s like they have given hands and legs to the LLMs:
- It has a great UI (yeah buggy as hell for some parts - but its better feeling than vscode)
- Huge catalogue of models for use + multimodal answers (unlike CC which limits us to one family)
- Code autocomplete is a nice addon
- Code indexing is done quite well, when I ask a question - it knows where to look
- The normal Agent mode can be used for everything - planning, coding, running browser things, running commands, debugging (although I miss custom modes - but I’ve started using User Rules in its place)
- Background agents - Run them before going to sleep, then in the morning, check the output (you can turn off your computer too)
I used default CC models before with a $20 subscription - it used to work great. But I don’t like CC because:
- I prefer an IDE over a terminal experience (terminals are limited to what a full-fledged UI can accomplish)
- My terminal gets hung from time to time, it keeps janking top to bottom (Windows CMD) - so I have to keep restarting it
- CC used to have different limits for the Opus 4 model, which used to get used up pretty quickly, like in an hour of just thinking
- It is a token hogger - I remapped CC’s models with my open router credentials and used GLM 4.7 as primary sonnet recently with haiku for summaries, it sucked up 3 dollars just to generate a CLAUDE.md file for my project (which has less than 5000 lines SLOC). Though GLM is cheaper than Claude, I did this to save money, but it backfired - I wonder if I did something wrong.
That said, this is a preference from a user experience perspective.
Cursor is great if you like GUI-based IDEs, CC might be good for you if you like working in terminals. Both tools have been getting better quite dramatically (CC was already better than Cursor v1; now Cursor is catching up).
Cost-wise, I’d look to other guys’ answers who pay WAY MORE than I do.
