Cursor vs Claude Code: we gave both the same coding jobs

Cursor and Claude Code are two of the best-known AI helpers for programmers. So we gave both the same three coding jobs from our secret test. To make it fair, we also let Cursor use the same AI model as Claude Code: Claude Sonnet 5.5. With the same model, both tools passed every secret check and were about as fast. The big difference was how much text they sent to the model: Claude Code read about 3.9 times as much per job.

What are Cursor and Claude Code?

Both are tools that let an AI write and change code for you. You give a job, like "fix this bug". The AI reads your files, makes changes and runs tests until it thinks the job is done.

• Claude Code is made by Anthropic. It works in a terminal, the text window programmers type commands in, and also in apps and code editors. It uses Anthropic's Claude models.
• Cursor is a code editor with AI built in, and it also has a command line tool. You can pick from many AI models, from Anthropic, OpenAI, Google and others, plus Cursor's own model, Composer.

Since August, Cursor belongs to SpaceX, Elon Musk's company. "Cursor has officially been acquired by SpaceX," Cursor wrote on its blog on 14 August 2026.

Our test: same jobs, four setups

We used the three jobs from our private coding test, the same ones as in our Claude Haiku 5.5 test. The tests in our public test kit are known by now, so these stay secret:

• D. Fix a money reader: code that turns "€ 1.234,56" into cents, with several bugs. 28 secret checks.
• E. Split a restaurant bill: shared dishes and a tip, with cents that must add up exactly. 17 checks.
• F. Read a CSV file: a spreadsheet saved as text, where commas, quotes and line breaks can hide inside a field. 17 checks.

Each tool got the job text and worked on its own: reading files, changing code and running tests. Nobody helped. Only after it was done did we run our secret checks. We tested four setups, each job three times:

1. Claude Code with Sonnet 5.5, effort "medium". 2. Cursor with Sonnet 5.5, also "medium". Same model, different tool. 3. Cursor with Composer 2.5, Cursor's own model. 4. Cursor with Auto, where Cursor picks the model. This is what you get for free.

Who did the better job?

With the same model, there was no winner. Claude Code and Cursor both passed all 186 secret checks with Sonnet 5.5, in every try.

Cursor's own model, Composer 2.5, passed 181 of 186. Auto passed 181 of 186, exactly the same. Their misses were small but real:

• Composer 2.5 twice rejected "1234,5", an amount without a thousands dot, and once "€ 0,5". In one CSV try it got two rules about line breaks wrong.
• Auto twice did not accept "12" as an amount, and twice rejected "€1.000", which means one thousand euros. Once it treated a special line-break character the wrong way.

So in our test, the model mattered more than the tool. Give Cursor and Claude Code the same model, and you get the same quality.

Which one was faster?

With Sonnet 5.5, Cursor was a little faster, but the difference was small. In the middle, one job took:

• Claude Code + Sonnet 5.5: 38 seconds.
• Cursor + Sonnet 5.5: 33 seconds.
• Cursor + Composer 2.5: 35 seconds.
• Cursor + Auto: 76 seconds.

Auto was the slowest by far: about twice as long as Sonnet 5.5 in either tool. The slowest Auto job took 98 seconds. Our Claude Code ran on a Windows PC and Cursor in a Linux sandbox. But the AI itself runs in the cloud either way, so we think that made little difference.

Same model, but Claude Code read much more

Here the tools really differed. Every time an AI tool takes a step, it sends the model your instructions, the files and everything done so far. That is counted in tokens, small pieces of text.

With the same model, Claude Code read about 360,000 tokens per job. Cursor read about 92,000. That is about 3.9 times less. Most of it is the same text, read again and again from a cheap cache, a kind of short-term memory.

At Anthropic's API prices, those tokens are worth about $0.25 per job in Claude Code and $0.11 in Cursor. You do not pay this per job with a subscription. But it shows that Cursor used the model more sparingly. One note: Cursor did not report any tokens written to the cache, while Claude Code did. If Cursor leaves those out, its real number is somewhat higher.

What does Cursor cost?

Cursor's price list on 10 October 2026:

• Hobby: free, with "limited Agent requests".
• Individual (Pro): $20 a month, with more use and "access to frontier models", the newest top models.
• Teams: $40 per person a month.
• Enterprise: price on request.

Claude Code has no separate price. It comes with Claude's paid plans, starting with Pro at $20 a month. We explain those plans and limits in our Claude Code pricing test. So for one person, both start at $20 a month.

What you can do on Cursor's free plan

We started our test with a free Cursor account. Cursor's command line tool refused to use Sonnet 5.5. The message said: "Free plans can only use Auto."

So on the free plan you cannot choose your model. Auto still did the jobs, with 181 of 186 checks passed. But it was the slowest setup, and you do not know which model did the work. Cursor's usage page only says "auto".

For Sonnet 5.5 and Composer 2.5, the owner of this site took Cursor Pro for $20.

What we saw on Cursor's usage page

Cursor shows every job on a usage page. On the free plan, it showed a price per job, even though we did not pay it. Our first two Auto jobs used about 100,000 tokens each. Cursor valued them at $0.07 and $0.10.

After the switch to Pro, the page only said "Included" for each job, with no price. That was the same for the Sonnet 5.5 jobs, which used about 83,000 to 116,000 tokens each there. That fits the numbers Cursor's command line tool gave us. For the whole day, the page counted about 860,000 tokens for our 9 Sonnet 5.5 jobs and about 1.1 million for the 9 Composer jobs. So with Pro you see how many tokens you use, but not what each job would cost.

Does SuperGrok include Cursor?

No. Many people may think so, because Cursor now belongs to SpaceX, just like Grok. And Cursor Pro now includes Grok Bot, the agent we wrote about in our Grok article.

But it does not work the other way around. The owner of this site has SuperGrok. Still, the Cursor account we tested with was on the free plan, and Cursor refused to use Sonnet 5.5. Cursor's help page confirms it: linking SuperGrok only gives you Grok Bot use in Cursor. "If you had no Cursor plan before linking, you still don't."

Which one should you pick?

In our test, the tool mattered less than the model. With Sonnet 5.5, both did perfect work in about the same time. So pick the tool that fits how you work:

• You already pay for Claude: Claude Code is included, and it did just as well. We compared it with OpenAI's tool in Codex vs Claude Code.
• You want an editor with AI built in, and many models to choose from: Cursor. With Pro you can use the same Sonnet 5.5.
• You want to pay nothing: Cursor's free plan works, but only with Auto, which was slower and made a few small mistakes.

This was a small test with three short jobs. Bigger projects, other languages and longer work may give other results.

How we checked

On 10 October 2026 we ran 36 coding jobs: 3 jobs, 3 tries, 4 setups. Claude Code ran on our Windows 11 PC with the site owner's Claude plan. Cursor's command line tool ran in a Linux sandbox at Vercel with a Cursor API key, as Cursor's docs describe. We did that because Windows Smart App Control blocked part of Cursor's tool on our PC. The secret checks never went into the sandbox: we copied the finished code back and checked it on our own PC.

The first 2 Auto jobs and our 2 failed tries with Sonnet 5.5 ran on the free plan. The failed tries do not count in the results. After that the site owner took Cursor Pro, so the rest of the Auto jobs and all Sonnet and Composer jobs ran on Pro. Times are measured from start to finish of each job. Token counts come from the tools themselves.

We read Cursor's price list, blog and help page on the same day. The screenshots are our own captures, cropped only. The usage page is from the account we tested with, cropped to our test rows. The charts were made by us. None of the images was made by AI.