** YOU MUST HAVE A CURRENT PAID CURSOR ACCOUNT IN ORDER TO USE ANY EXTERNAL LLM
For heavy development, using your own Anthropic API key will definitely be cheaper.
Here’s why:
Cursor’s Pricing:
API pricing + 20% markup when you use their system
Even with Cursor Pro, you’ll likely exceed the 500 “fast” requests quickly with heavy development
Direct Anthropic API:
No markup - you pay Anthropic’s rates directly
Claude Sonnet 4 pricing: you pay only for tokens used
Much more cost-effective for high volume
The Math: If you’re doing heavy development and making hundreds of requests daily, that 20% markup from Cursor adds up fast. Plus, once you exceed Cursor’s included usage, you’re paying the markup on every additional request.
My Recommendation for Heavy Development:
Get an Anthropic API account directly
Set up billing with Anthropic
Use your API key in Cursor
You can still keep a basic Cursor subscription for other features, but use your own API key for Claude
You’ll save significantly on costs while getting the same Claude Sonnet 4 performance. The setup is straightforward - just add your Anthropic API key in Cursor’s settings.
Here’s how to set up Claude Sonnet 4 with Cursor using your own API key:
Step 3: Start Using It
Use Cursor’s chat feature or inline suggestions
Sonnet 4 should now be available in the model dropdown
You’ll be billed directly by Anthropic based on usage
Quick Tip: Set up billing alerts in your Anthropic console so you can monitor your usage, especially during heavy development.
The exact menu locations might vary slightly depending on your Cursor version, but look for “Models,” “AI,” or “API Keys” in the settings. Most users find it pretty intuitive once they’re in the settings.
Need help with any specific step, or running into issues finding the right settings menu?
But if we do that, do we use Claude directly or do we still use Cursor’s optimized version?
For example when Claude 3.7 was released it worked horrible with cursor until they made some modifications.
Then when Claude 4 was released Cursor posted this on twitter: We’ve been very impressed by its coding ability. It is much easier to control than 3.7 and is excellent at understanding codebases.
So it seems that they might be doing some behind the scenes things with Claude to make it work better with Cursor. I wonder if we still use that version or does it just connect to the claude API and use it directly without their modifications?
Also for Claudes API, do we subscribe to one of those plans then we get a certain amount of API usage? Or is it pay as you go?
Or, install Claude Code in the command line, open Cursor, open a terminal inside Cursor, type «claude» (Enter), and Claude will install its own plugin into Cursor automatically. I mostly stopped Cursor’s AI integration with Claude - it’s more expensive, way slower, and less precise.
I don’t realize how much it costs, having to have the $100 package to do this ?
I think it’s a real shame that the cursor team didn’t offer another package with less clamping, more context etc… Because I love Claude 4 and, frankly, the pro version of cursor doesn’t let me enjoy it properly.
Cursor offered 2 days of Sonnet 4 for a test run. I took total advantage of that. The experience was awesome. So awesome I went and bought the Claude Max Plan. I’ve not installed my API yet, but will be testing it today.
I will post my findings here.
This makes me wonder if there is a difference in using Terminal “<>” and inputting the API key into the SETTINGS window?
Thanks for the heads up.I’m going to try both and see if there is a difference.
Sessions vary with task complexity as well. It takes hours to know how a model really performs.
I’m really interested too. Keep us posted. Does this mean we get the max mode in “unlimited” we pay no extra fees?
I’d like as much information as possible. The best implementation on cursor
For me, Sonnet 4 has been pathetic and messed up my landing page and couldn’t even fix it after many iterations despite telling it where the issue is. Somehow, it made my NavBar a vertical column instead of a horizontal top bar with h-16. It was using h-full where it isn’t even needed and I explicitly mentioned that but still couldn’t fix. I was simply laughing at it being a SOTA model for coding. Yes, it was provided with PRD and Frontend implementation instructions. In short, I fixed it myself.
Problem ist, currently this comes with two crucial issues :
tool call sequences are often interrupted with API rate exhaustion errors - seems to happen way too often and early in the process
Cursor’s internal model management seems to react to that in weird ways. Now SOnnet 4 disappeared from the model selection entirely ( note - this is having run out of my monthly quota and switched to API for a few hours - API rate exhaustions kept piling up, then the model disappeared completely - IDE restarts do not solve that - heck even disabling API again won’t bring it back.
Quite annoying and thinking to go for Claude Max and Claude Code integration in cursor
you can pick the models in the settings which may be why the model disappeared. the issue @xblabs mentioned about rate limits is real.
Cursor needs to fix their pricing or people will bail because charging people multiples of what the backend cost is not a viable solution for heavy users. I find myself switch to copilot in another IDE to do work from time to time simply to avoid getting mugged.
Hey all, totally appreciate the desire to use Cursor with an API key to offset some of the costs.
Out most productive users often lean into the various models we offer more, and want the ability to use multiple models from different providers without needing to create and manage a bunch of API keys.
In addition, newer Cursor features will continue to involve custom models and features that need to be charged for outside of the main model you pick. This is already the case for Tab and Apply, which both use custom models to do their jobs.
Think of Cursor like an ice cream shop - you could definitely go and buy the same or similar ice cream yourself in a big tub from a shop, serve it in a bowl and save some money. But then you’d miss out on the cone, the sauce and the convenience of being able to pick from any flavour you want every time, without having to buy every flavour yourself!
Happy to field any questions anyone may have on this!
Thank you for your reply.
I have a real question, why don’t you make a more expensive plan offering the maximum context of the model used? With a certain number of requests per hour, or an unlimited system or something. Because for those with a lot of codebase and a big database, the slow mode and some templates are not necessarily enough anymore. Personally, I love Claude 4 and it’s frustrating not to be able to use it the way you want. You could easily make an offer between $60 and $90.
Is this planned? Many of us have been asking for this.
If you have a complex piece you need Claude Sonnet-4 to do, just pay the $17 for Claude. Now you have use of Sonnet-4.
Next, give Clause a High Level explanation of your project, then give claude detailed instructions on the features you want to add. You will see a visual of what cluade makes, don’t be concerned about what the UI looks like, as you will be only using the backend funcitonality.
NEXT →
In your project create a directory “documentation”.
Copy and paste your Sonnet-4 code into a .md file (filename.md) .
Now Tell Cursor to review the “documenatation” folder and view the file you made with the code in it.
Explain to Cursor that this is the functionality you want to add to your project.
