I’ve been using Cursor for 1 week now and absolutely love it! I primarily use the Composer feature as it generates amazing responses when it has the context of multiple files. I upgraded my amount of fast premium requests to 1000 per month. Honestly, I would like to have 10,000-15,000 fast premium requests per month as the larger models perform much better on non-trivial tasks. I understand that the larger models are more expensive to inference, but I would appreciate if the Cursor team could work on decreasing the cost of a fast premium request. Right now, the cost is .04 USD per request. Getting the cost closer to .02 USD per request would be amazing! Thank you!
3 Likes
What on earth are you doing that you need 15,000 requests to an LLM in a month in a non-automated context? That’s more than 1 per minute every minute for 4 weeks at 50hrs a week. That’s not even sufficient time to author reasonable prompts, much less read the output. Are you just YOLO “coding” with zero understanding of what’s happening? Genuinely perplexed and curious.
2 Likes
So, I’m not actually using that many requests. I just used those numbers as a rough upper limit to how many requests I could use in a month (sure, my estimate may have been a little high). I am often iterating quickly with the LLM to fix bugs, and sometimes without reviewing the code until the bug is fixed. I realize this is a bit lazy and possibly risky. However, it is very efficient and I review any files that were modified during the process of fixing the bug before committing. Overall, the larger models are less brittle to work with than the smaller models, and if the cost were the same I would prefer to use the larger models for most tasks.
In my experience this is a rather quick path to an unmaintainable app, but if it’s working for you, maybe using your own API keys is the move. I do and it’s cheaper than paying for Cursor, but it sounds like your usage is dramatically higher.
Thank you for the suggestion on using my own API keys, I will look into that. I agree that you can quickly get yourself in trouble by having the LLM generate a bunch of responses without reviewing each one. In my case, I’m committing my code every time that I review it and test that everything is working. When things go awry I hard reset to the last working version. I realize this approach is a bit hacky, but it does help with rapid iteration. Thank you for the feedback.