Hello,
I use both tools in a 100% “vibe-coding” workflow, because I don’t know Python syntax. I do know how to code in C and C++ though, but today most APIs and libraries are in Python, so I have to use this language as well.
I mainly develop scientific applications, typically between 3 000 and 10 000 lines of code. My projects involve a lot of numerical computation: integrals, interpolation, and geometric calculations.
Since Cursor is older than Claude Code, most of the hype on YouTube and social networks seems to suggest that Claude Code is much better than Cursor. In my experience, this is completely wrong, and I’ll explain why. This is more based on practical usage than on formal benchmarks, so I may be wrong, but here is what I’ve observed.
1) Speed and model choice
Cursor is faster and gives easy access to very fast models for simple tasks, which is extremely useful for simple tasks.
2) Easy switching between models
Cursor makes it very easy to switch between small models for simple tasks and large models for difficult ones.
For example, for graphics-related problems, the only LLM that consistently solved my issues was GPT-5.2. Both Opus 4.5 and Sonnet 4.5 failed many times. For simpler tasks, I use lighter models. I don’t know whether Claude Code allows model switching, but by default it only provides Anthropic models.
3) Cost
Cursor also feels much cheaper to me. You can add DeepSeek, and you even have access to Grok, which is currently free.
My experience is that it is a great tools but we still needs a lot of time to control and correct what the AI does. I found dozens of mistakes in the code, even if I try to specify everything : conversion problems, integral bounds, confusion between times (UTC, local time…), matplotlib difficulties to put the legends at the good places, forget to do interpolation when needed and asked, forget to define general functions that are used many times even if asked…
There are some drawback in Cursor : I think that the models we use are not clearely defined. We don’t know what there are, if there are fine tuned or not, where they work…
The first which would give two options might win the game :
1/ possibility to run on personnal working station with small models. Even though personnal GPU might not have enough power, we have to learn to code with smaller models otherwise it would be too expensive. I would love to mount my own system, based on solar energy during the day for instance in order to reduce my carbon costs using AI. This is something really promissing.
2/ clarify the models and tokens we use for each task or questions. Especially smaller ones in order to reduce the costs and CO2 emission. We have to improve ourselves to know if using a small models for a task would work or not. We need metrics to estimate everything because we have along with LLM and agents to improve ourselves as well and don’t do everything in a black box
3/ options to limitate what can be down with application like Cursor and what we can do on our own. We use too many tokens to do something which is not really valuable with AI. For instance, if we break up the code in small parts and small files, that may reduce the tokens used.