A Review of Using Both Cursor and Claude Code

These days, it seems like many people only use Claude Code, and I haven't seen anyone subscribing to and using Cursor.

A while ago, I saw positive comments praising the excellent performance of Cursor's new proprietary model, Composer 2. Out of curiosity, I subscribed to the $20 Cursor Pro about two weeks ago, so I am currently subscribed to all three: Claude Pro, Cursor Pro, and ChatGPT Plus.

Looking at posts written by others, most said that Claude Code is much better than Cursor, so I wondered if that was really true. To give you the conclusion first, it is true.

There are various reasons, but whatever the reason, I want to say that using Claude Code is a better choice than using Cursor.

If you only use AI coding agents once every few days or if you write code yourself more often, section 1 of this post won't apply to you. If you are such a person, please read from section 2.

1. Price and Usage Limits

Since I am still a student, I cannot afford to pay high subscription fees like $100 or $200. Paying $20 a month right now is tight, so Claude Max or Cursor Ultra are completely out of the question.

I don't trust AI-generated code and make it check for bugs or security issues multiple times, which is another reason I use a lot of tokens. However, without such verification, I don't know what problems might occur when deploying to production, so I can't just skip it.

Then what about Claude Pro (Claude Code)?

People often say that Pro is just a trial version of Max. It's not actually to that extent, but it's true that the limits are very tight.

I only use Opus on the Claude.ai website and only use Sonnet in Claude Code, yet when I finish building and testing a single feature with Claude Code, the 5-hour limit is almost full, and the 7-day limit is also filled by about 4-6%.

To give you a tip, if there is a 7-day limit like Claude or ChatGPT, using about 14% a day, and if there is a 30-day limit like OpenCode Go, Copilot, or Cursor, using about 3.2% a day will allow you to use it all the way to the usage limit reset date without getting blocked.
I use a program called OpenUsage, which is convenient because it automatically does the calculations mentioned above.

Still, since it's an official subscription that directly provides the Claude model, it seems to give more usage than third parties like Cursor that use the API.

Especially when compared to Cursor, which I will explain below, you can see that this is a godsend.

However, Cursor Pro is even tighter.

Based on the Pro plan, Cursor gives $20 worth of API usage and an estimated $80 worth of Auto+Composer usage. The total is $100, which seems like a lot, but it's not.

First, I should explain these two pools. The API usage pool is deducted when calling third-party models like Claude or Codex within Cursor. The Auto+Composer pool is deducted when using Auto mode or Cursor's proprietary models, Composer 1.5 or 2.

I will explain more detailed information about the Composer model below.

So ultimately, if you subscribe to Cursor Pro, you can only use a measly $20 worth of high-performance third-party models, and you can use a lot of their slightly lower-performing proprietary models. But in fact, even if you only use the proprietary models, it can be insufficient.

At first, when using the Composer 2 model, it was fast, the performance wasn't unusable, and the usage limit didn't fill up much, so I used it in Fast mode, which uses 3 times more usage.

But from about the 3rd day, I saw the usage limit filling up faster and faster, and now I end up using way past the safe usage amount.

I found out that for the first 2 days after the model was released, they had a double usage event so I could use more, and after this ended, it returned to the normal usage limit, making it insufficient.

If your budget is small, the official Claude is definitely better.

I think Claude is a better choice in this regard.

My plan is to cancel my Cursor subscription when it ends and only use the Claude Pro subscription.

2. Model and Harness (Shell) Performance

Claude Code can use the Haiku, Sonnet, and Opus models.

Cursor can use most models including Claude models, Codex (GPT), Kimi, and Gemini, and also provides its proprietary model, Composer.

I only used Sonnet 4.6 in Claude Code, and in Cursor, I used Composer 2 as the main and Sonnet 4.6 as an auxiliary.

Even using the same Sonnet model yields different results.

Claude Code gives the model more freedom and allows it to do various things, whereas Cursor gives more authority to the developer and is designed with slightly more restrictions on the model.

For example, if I give it a task unrelated to a coding project, like one of the things I asked today: "I'm using windscribe port forwarding, but open webui can't be accessed from the outside. Please fix this issue," Claude Code moves to an external directory to work and even manipulates other apps, but Cursor replies that it cannot perform the task because it cannot be done in this working directory.

On the other hand, if it's inside a project, it's the opposite. Claude Code requires approval to execute unless explicitly allowed, but Cursor's default setting is to apply and execute immediately without approval if it's a task within the project.

Perhaps because of this, when the work is done, Claude Code goes to external directories where other related projects are located to perform verification and testing, but Cursor only does some formatting unless explicitly requested, and doesn't do actual testing, creating the annoyance of having to test it myself.

And the difference between testing it yourself and the model finishing the testing is bigger than you'd think. When the model tests, it immediately knows the output or error details and patches it, but when a human tests, they have to explain it in words, the model modifies it, and if it doesn't work, it's reverted, and you have to explain it in more detail again, which repeats.

I will raise Claude Code's hand in this aspect as well.

Cursor's proprietary model is usable.

Cursor also develops its own proprietary models. And this is also the reason I subscribed to Cursor.

The Composer 2 model was released a few weeks ago, and this model is based on and improved from (smarter than Kimi K2.5) the Kimi K2.5 model by China's Moonshot AI.

The biggest features of this model are that it is fast by default, and there is a Fast mode that can make it even faster (but uses 3 times more usage).

If you use Fast mode, it is exceptionally fast, modifying one file per 0.2 seconds when commanded. This is much faster than Claude Code's fastest model, Haiku, while its performance is slightly below Sonnet, so it is commendable.

Looking at pure model performance, it modifies code well, but its lack of language skills leaves something to be desired in actual use. Claude is a large model itself and has learned a lot about language, so even if you speak in Korean, it understands like a human, builds well, and explains well. However, Composer is based on China's Kimi and underwent additional training mainly with English data, so if you speak casually, it often understands less well than when speaking to Claude.

I thought it would be fine if I just explained it clearly, but it's more annoying than I thought.

What Claude does when you just say "~ doesn't work in ~", you have to explain to Composer like "When I try to do ~, a ~ window pops up, a ~ error appears in the console, and as a result, there is a problem where ~ doesn't work, so please check and fix it." This is quite a hassle.

In conclusion, the Composer 2 model is usable, but you have to explain well for it to build well.

If you are confident in explaining things clearly, it will be a good choice at a cheap price. But since I couldn't do that, I have some regrets.

3. Interface Usability

This is inevitably more convenient with the GUI-based Cursor.

Claude Code can also be used with a GUI if you use the Claude program or the Claude Code extension, but there is a very big problem that its features are severely limited and you cannot use most of Claude Code's unique features.

When Claude Code loads a session, the conversation at the top gets cut off. This is unavoidable because it is terminal-based. And there is also the problem that when the context is compressed, all previous conversations are blown away.

Cursor is designed to load more when you scroll up, reducing the load and making it easier to view existing conversations.

Besides that, there are minor differences, such as when copying, Claude Code copies the entire UI, whereas Cursor, being a GUI, naturally only copies the selected part.

I think we'll have to wait and see about this part, but currently, Claude Code's development roadmap includes a TUI update scheduled, which will make it more convenient to use, like OpenCode, where it's a terminal but you can use a mouse like a GUI and copy only selected parts.

It will probably become heavier than before, but if it becomes easier to use, it's welcome.

In section 3, I will side with Cursor.

This is an unavoidable choice since it's GUI vs CLI in the first place.

Of course, using a terminal has its conveniences, but for most people, working with a GUI will be more familiar and comfortable.

4. So what should you use?

If you only have a budget of about $20, I recommend Claude Pro more because it gives more usage.

If your budget goes up to $100 or $200, I recommend trying both and deciding. You might have different thoughts or tendencies than what I felt.

If you have no real development experience and have never used a terminal (a so-called vibe coder), Cursor would be a better choice, but in this case, rather than Claude Code or Cursor, I recommend using other products like Antigravity where prototypes come out faster and usage limits are more generous.

Whether it's Claude Code or Cursor, they are products made considering that the user has prior development experience, so they are bound to be difficult. On the other hand, Antigravity feels more like it was made for users with no development experience.

5. Try PrivateStater

There is a service I run called PrivateStater.

Most people reading this post probably have their own website, and it's a service you can use well in that case.

It provides analytics, captcha, and feedback widgets. For more details, please refer to the PrivateStater website.