Cursor Ate My Token Budget. Claude Code Gave It Back.

However, I have one major issue with it. This issue is so confusing and frustrating that it is forcing me to look into other options.

This problem ultimately boils down to ‘value’. When you part with your money, you want to make sure you’re getting ‘value’. I feel with Cursor I am not getting as much ‘value’ as I could be getting.

Let’s dive into the specifics, see where the issue is and then I will show you some hopefuly alternatives I have been exploring.

The Major Issue With Cursor

It’s all to do with how many tokens Cursor seems to be burning through whenever I use it.

Over the past month I have been working on a personal project to really explore using AI heavily when writing code. I have heard enough hype and opinions that I realised it was time to give AI generated code a good go.

I currently use Cursor as my main editor (although this shifts quite regularly at the moment) and have been using primarily the Opus 4.5/Sonnet 4.6 and Composer 1.5 models.

My general approach is if I am building out a new feature I will go into “Plan mode” with one of the Opus models and use that model to design out the feature. Then I would get the same model to implement the majority of the code.

If I am working on small UI improvements, writing tests or small updates to logic I will try to use Composer 1.5 first and if it’s not up to the job I will switch over to Opus.

It didn’t take long, especially with the Opus models, to reach my limit. Once I had reached this limit I couldn’t even use Composer anymore. After some deliberation I soon realised I had no choice but to update to the $60 subscription. (Obviously I also had the option to continue the project without using AI but come on! It’s 2026… if you’re not using AI then you’re a fool. Right?).

This is when I started looking into how Cursor counts usage. I was shocked.

How Cursor Splits Up Usage

Cursor splits up your usage into two buckets:

• Auto + Composer
• API (If you use Claude Opus 4.5 or Codex 5.3 for example)

Press enter or click to view image in full sizeAPI usage quickly surged up to 99%

As you can see, even with the $60, I was soon up in the 90% usage and I am now at 99% and still have some days left until my subscription renews. I am stuck using Composer or Auto.

This API usage seemed to go up really fast. I was making use of the models but I didn’t think I was doing anything excessive. I wasn’t letting it run neverending loops. My prompts were very precise in terms of what files to look at and what changes I wanted. I therefore went and looked into the “Usage”. Again, I was shocked.

Press enter or click to view image in full sizeRidiculous token usage like 4.4M or 1.4M

I was seeing some outrageous numbers — 4.4M or 1.4M tokens!! This was Composer which I only use for relatively basic frontend tasks — “Create a component to show this specific data”. “Refactor this component out into smaller reusable components” — Nothing logically difficult, just tasks I didn’t want to bother coding myself.

The usage on the “Opus 4.5” model was of similar crazy numbers.

Press enter or click to view image in full size2.4M or 4.4M tokens for one usage.

What was going on here?

Once I realised how I was burning through tokens with Cursor I started looking online for reasons for this. Maybe I was doing something stupid. Maybe I had some setting wrong. I didn’t find anything obvious, just a bunch of tips about renewing sessions regularly, keeping requests small and focused, not using “max” models except for really tricky things.

I had to find a better way.

Claude Code To The Rescue

I happen also to subscribe to Claude.ai for $20 a month. I realised upon investigations that included in this price is Claude code usage. I had been avoiding this tool for a while, as the Cursor experience was really good, but now I was forced into trying it.

I downloaded the CLI tool.

curl -fsSL https://claude.ai/install.sh | bash

The one thing I noticed straight away on the Claude code CLI was it showed you the token usage as it worked on problems. I immediately noticed that the token usage in Claude code was much lower.

Even the more complex tasks with planning and code production were using around 100k — 200k tokens for the whole feature.

Claude code is apparently the most popular AI Tool for developers at the moment.

Press enter or click to view image in full sizeI did take this from the Pragmatic Engineer substack email. I really recommend his emails. Very informative.

But after using it for a few weeks I have to say I wasn’t overly impressed with the UI/UX. I mean it’s fine and works but there are a few key things that I think Cursor does much better.

• Easily able to select specific lines of code to give as context to the AI chat.
• Showing the thinking logic and commands it is running as it’s doing its thing.
• More obvious UI in terms of what code has changed in which files.

Claude Code Isn’t Just a CLI Tool

I just so happened to demo this tool at work as many team mates had not actually used the tool before. Afterwards, one of my colleagues, after hearing me complain about the UX of the CLI tool, asked if I had tried the VS-Code/Cursor extension.

I didn’t even know there was such a thing!

I tried it and instantly saw the light.

The experience was much more aligned to the Cursor experience. A different shortcut to open it and it’s not quite as “native” in terms of its feel but it’s close enough.

I started using it in a similar manner to how I was using Cursor. However, this time when checking my usage the numbers were much more aligned to my expectations. No longer was I burning through tokens to complete relatively simple tasks.

This is definitely the way.

Conclusion

Since I have discovered this Claude code extension I have been using it 95% of the time. I have switched my Cursor subscription back to $20 and honestly I barely use it now.

I still only have my $20 Claude subscription and this seems to be enough for the amount of AI usage I need for both coding and what I ask it on the Web browser.

My colleague actually has taken it one step further and uses the Codex extension too in VSCode. He basically pays $20 for Claude and $20 for Chat GPT and nothing for Cursor. He then gets the Cursor experience but access to the latest models at a much fairer price.

This is where it gets back to “Value”. Ultimately with this approach you’re getting better value as you’re paying less for the same service. Which leads to an interesting point. Ultimately the companies that own the models are the ones in control here. They want you to go through them directly and that it where the best value lies.

Cursor currently have the best UX/UI and tooling around but their pricing model is heavily dictated by the AI companies. They also built their whole empire on top of VSCode. This whole situation seems precarious at best.

Who knows what is going to happen. Humans do tend towards the easiest and most convenient option and current Cursor is that option. However, I have realised through this process it certainly isn’t the best ‘value’ on the market. It pays to shop around.

Subscribe to My Weekly Updates on Medium!

Enjoyed This Post?

If you found this blog post helpful, why not stay updated with my latest content? Subscribe to receive email notifications every time I publish.

If you're feeling really generous you can buy me a coffee. (Btw, I really like coffee…)

What You'll Get

• Exciting Discoveries: Be the first to know about the latest tools and libraries.
• How-To Guides: Step-by-step articles to enhance your development skills.
• Opinion Pieces: Thought-provoking insights into the world of frontend development.

Join Our Community

I live in the vibrant city of Prague, Czech Republic, with my family. My blog is more than just articles; it's a community of like-minded developers who share a love for innovation and learning.

About me

I'm a passionate Frontend Developer specialising in React and TypeScript. My professional journey revolves around exploring and mastering new tools and libraries within the JavaScript ecosystem.

Check out my LinkedIn and Github if you are interested.