Introduction
My journey into AI-assisted coding began some time ago, but my last three months have been a deep dive into the current landscape of tools and models. My first experience was with GPT-3, a little after it was released. Since then, a lot has changed and different models have been introduced, each of them with powerful features.
To put these tools to the test, I’ve spent the last few months working on a diverse set of projects: solo-developing MQL4/MQL5 trading bots, collaborating on a Python-based auto-trading system, and building out my personal Data Science portfolio.
My toolkit for this period included premium subscriptions to Github Copilot Pro and Gemini Pro, a trial (and later, the free version) of Cursor, and free-tier access to the Claude and ChatGPT chatbots.
Comparing Platforms
Cursor IDE
I had the best experience with Cursor, especially when I was on the free trial period. Its model auto-selection is superb and most of the time the results were beyond my expectations. Since I had a Github Copilot Pro subscription and my Cursor free trial ended, I decided to switch back to Github Copilot. Cursor’s free version is also good, but not for complex projects and tasks.
GitHub Copilot Pro
GitHub Copilot Pro was the second-best. I couldn’t find anything similar to Cursor’s auto model selection, so I had to find the best options through trial and error. I will mention each model’s performance in the next section.
A standout feature for both Cursor and Github Copilot is their ability to automatically generate commit messages. This was a huge time-saver in my daily workflow. While the AI-generated messages could sometimes be overly verbose for commits with many changes, I found myself approving them over 80% of the time—they were often more descriptive than what I would have written myself.
I primarily used these AI IDEs in ‘agent mode,’ which I found more efficient than the traditional chat-and-copy-paste workflow. This approach lets you approve or decline any part of the agentic changes directly in the code, giving you full control. It’s more convenient than asking the AI and making the changes yourself, and any necessary explanations are still included in the AI response anyway.
AI Chatbots
For the last part of this section, I want to mention chatbots. Most of my time was spent in AI IDEs, but sometimes I used chatbots as well, especially when I wasn’t satisfied with the IDE’s result or changes and I couldn’t get what I was looking for. In terms of platforms, I preferred Claude and Gemini over ChatGPT. But a platform is only as good as the models running on it, which brings me to my next point.
Comparing Models
Choosing the right model is arguably the most critical—and challenging—part of using these AI tools. It’s the difference between a seamless ‘vibe coding‘ session and a frustrating one. There’s no absolute answer for all cases. Also, sometimes it comes down to developers’ personal preferences. But I want to share my point of view.
Gemini 2.5
If I could only use one model for everything, my top pick would be Gemini 2.5. It can handle complex tasks, understand what exactly you want and, more importantly, it does as much as you ask, no more or less. It works surprisingly well in long conversations, and I rarely find the need to restart the conversation. It works great in both IDEs and its chatbot platform, and I used it a lot.
Claude’s Sonnet 4
Coming in a very close second is Claude’s Sonnet 4. While nearly on par with Gemini 2.5, it has a distinct personality. In almost every case, it gave me way more than I needed. It may sound like a benefit, but trust me, it usually becomes annoying. For example, I ask for a simple correction in code, and I end up with two extra test modules that the model decided to create to make sure its changes work as it should, and I had to delete those files and remember to ask it strictly to do exactly what the task is, nothing more or less.
However, it sometimes understands complex challenges better than Gemini 2.5, as long as you don’t continue the conversation for too long. I use Sonnet 4 for complex issues, and most of the time I don’t request code changes. I get its advice on algorithms, improvements, etc., and try to implement them myself (or with Gemini’s help!) For small and easy tasks, I never use Sonnet 4 since it complicates things.
GPT 4.1/5
Finally, I used the GPT models (4.1 and the recent 5) primarily as a specialist for specific tasks. I use GPT models for small tasks, or when I’m sure that there is not much complexity involved. For example, for refactoring a large part of code or something like that. Its speed is usually higher than the other models. But one annoying thing that I encountered was some basic mistakes, such as wrong indentation in Python code. It only happens when I use GPT models in Cursor or Github Copilot, and it doesn’t happen in ChatGPT itself. I’m still not sure what’s the real cause of this, but in some cases, these agentic changes by GPT wasted my time and I had to debug the code to find the indentation issues. There were other models that I could choose such as Grok, but I didn’t use them a lot so I can’t comment on their performance.
AI Model Comparison for Coding
| Feature | Gemini 2.5 | Claude’s Sonnet 4 | GPT Models (4.1 & 5) |
| Overall Rating | The author’s top pick for an all-around model. | A very close second, with a distinct personality. | A specialist for specific tasks. |
| Best Use Case | Versatile for both complex and simple tasks in IDEs and chatbots. | Getting high-level advice on complex algorithms and improvements. Not recommended for simple tasks. | Small, low-complexity tasks, or specific jobs like refactoring a large part of the code. |
| Key Strengths | Understands requests precisely and delivers exactly what is asked—no more, no less. Performs very well in long conversations without needing a restart. | Sometimes understands complex challenges better than Gemini. | Its speed is usually higher than the other models. |
| Potential Downsides | No significant downsides were mentioned in the text. | Frequently provides much more than requested, which can be annoying. May add extra files like test modules without being asked. Not ideal for long, continuous conversations. | Can make basic mistakes like wrong indentation in Python code. This indentation issue specifically happened when using GPT models inside IDEs like Cursor or GitHub Copilot, not in ChatGPT itself. |
Final Thoughts and Takeaways
Overall, my three-month deep dive into AI-assisted coding was overwhelmingly positive. It helped me a lot in solving problems, speeding up my coding, and correcting my bugs. There were times when I asked for a change and hit enter and while waiting for the response, I saw another issue in my code. I thought to myself, “Remember to fix that after the AI changes come in.” And then I saw the AI had solved that too!
However, getting the most out of these tools requires the right mindset. Here are my key takeaways:
- Master Prompt Engineering: Familiarize yourself with prompting techniques to get the best results.
- Trust but Verify: Never trust the output 100%. Checking the AI’s work is a must.
- Work Incrementally: The best results come from using the AI on smaller, focused tasks, step-by-step.
- Stay in Command: It’s not wise to put the AI in autopilot mode (yet!). Think of it as a powerful assistant that enhances your experience, not as a replacement for your judgment.
- Keep Your Skills Sharp: Intentionally handle tasks manually from time to time, even repetitive ones. This practice ensures your fundamental skills stay honed and prevents the AI from becoming a crutch instead of a tool.
