Claude vs Deepseek vs ChatGPT vs Grok vs Google Gemini

I have been building various tools using all 3 of these LLMs. Deepseek R1 is really useful when you need to figure out the thought process behind a suggestion. I have been building a stock predicter app, a rss tp blue sky poster, this blog.

This blog has probably been the most complicated for some reason. The process to get gatsby to produce the static files from markdown is unexplainably difficult. It would have been easier to build this blog manually with React and HTML.

ChatGPT

I think ChatGPT is probably the best of the three when trying to create NodeJS applications, but its free tier runs out fairly quickly. Once the free tier is used up, I go to Claude, and finally DeepSeek. DeepSeek is still very good at solving problems, even better when R1 is enabled, but it can sometimes produce code that has the same bug it in over and over again.

I wanted to try using agents inside VS Code, but I keep hitting my free tier before I get a chance to use it in my IDE. I've used copilot a lot, but I enjoy coding sometimes so I just find myself coding it out individually.

Its very good at technical approaches, such as discussing an approach before implementation. It isn't afraid to just cut to the chase, which the other LLMs need to be reminded of occaisionally.

Deepseek

Deepseek is an excellent LLM. There have been reports that Deepseek does have centers that are directly accessible by the Chinese government. While concerns are warranted, it seems to me that Google / Meta (American companies) also share data with the American government. Them reporting that I am building a personal iOS app doesn't seem like sensitive information. Data Privacy is obviously very important, but to be able to use an LLM for free, that has a comparable coding capability to ChatGPT Plus (which costs $20 a month) seems like its a fair trade off.

Experience with Grok

Grok is extremely verbose, so its good at talking through problems. It doesn't seem very good at actually fixing coding problems. I had a hard time coding a layout in SwiftUI and grok seemed to understand the problem, but multiple times produced code that had the problem I described. Eventually I had to just figure out the problem myself using hints about what it was talking about.

It seemed like it had enough knowledge, but didn't know how to apply it correctly. The responses are often erratic and unorganized.

Claude 3.7 Sonnet

Claude has a cool split screen view, in which it will modify code on the right hand side (similar to Groks). I'm sure other tools do this, but it seemed to understand what we were doing and split the screen immediately. It also "versions" the code on the right (like v2) to indicate something changed.

I have found that its web interface can become very slow in Firefox, it will frequently say "This page is slowing down Firefox, To speed up close this page".

Claude Has a great user experience when editing multiple files. It will say here is the first edited file and then the second edited file and so on, makes it much easier when needing to modify multiple files as is often the case in complicated Xcode swift projects.

After working with this tool for hours, Claude will often produce code that contains errors. The same prompt will cause DeepSeek to produce similar code, but simpler and without errors.

Google Gemini 2.0

Google Gemini 2.0 is probably one of the worst LLMs to help with code. Working in a python project, the problems were clear that we needed to install (using !pip install), but it failed to even recognize some of the simplest errors. 2.5 Pro experimental was better, but the project had dependency issues right out of the box.

One thing that was a positive, was I was able to open the Python file directly in Google Colab.

Google Gemini 2.5 Flash / Pro

For some reason, this model is lauded by many folks in the industry. My experience has not been great. I tried creating a function for a SwiftUI project that tried to figure out what image model was best for the current device. It created code with errors every single time. I kept giving it the errrors and it would respond with the previous version of the code.

Overall

I think Deepseek R1 is the best LLM to use when coding small to medium projects. ChatGPT and Claude 3.7 tied as runners up.