My Review of AI Coding IDEs: Real-World Impressions
Part I (Cursor, Codex and Claude)
I have been using AI coding tools for about a year now, my first experience was with Cline, and then when Cursor came out it ended up winning me over with the completely AI integrated experience it offered as opposed to the partial extension based experience Cline offered at the time in VS Code. With that said, I am a bit of an agnostic when it comes to AI assisted coding workflows.
As of today, I alternate between Cursor, which is always my go to and most comfortable and preferred IDE to Trae, Windsurf, Claude Code, Codex and until very recently AntiGravity. Being a former front end developer, I have a natural preference for IDEs because of the 360 view it gives me of code, configuration, integrations and operations. Terminals I found useful, but until Claude Code and Codex, I would have never dreamed of coding primarily in a terminal, apologies to all the vim lovers out there. This is my retour d’experience ( literally feedback from experience, french term for giving feedback after actually experiencing something) on the AI Coding tools I have used and keep using over the past year.
Before getting in the reviews, let’s establish some baseline facts:
I alternate Claude Code, Codex, Trae, Windsurf and Antigravity for personal projects and pay out of my own pocket for the Pro plans for each service or fund API Key usage. The point is I have to be more price conscious because it’s my personal budget.
I use Cursor at work, and we have a reasonable budget for it, less than what I would like but reasonable enough.
This is not mean as a how to deep technical review, but more of a “vibe” as in how well does this tool gel with my preferred workflow review.
Cursor
King of the hill. The reason why I very gradually but quickly abandoned Cline was because Cursor went way further in making the AI assisted coding central to the IDE experience. What won me over with Cursor and that continues to fuel my preference up to today is that you can really feel in the design that this is a tool that was built by developers for other developers (I don’t know if that is true or not, I could find out but who has time, what matters is how people make you feel). There is that element of love in it, where after seeing a feature for a first time for yourself in action, you’re like these people get me. Developer centric features like Plan mode, Agent Review, and my own favorite, BugBot really resonate with me and highlight that they understand my pain points and have built a custom solution to address them. This is the reason why it’s my main driver. Performance wide it’s been okay, stable, it’s still VS Code but my M3 is still holding up for the most part. Now pricing is where Cursor is starting to lose me a bit, over the past year, it’s gotten increasingly expensive and I only use it for work, because work has the budget, to a degree. I would have loved to use Bugbot at the team level; I have found it to be an effective and insightful code review agent but the current cost is not justifiable in an org still reeling from sticker shock at AI prices. We found ways around but the simpleness of the Bugbot integration was nice. Cursor is nice, really effective, but only if you have the budget in my opinion, that’s why I only use it for work.
Claude Code
The French man is third in my list purely for economical reason as of today. I am too broke for Claude.
This is how I originally started this paragraph in the first draft of this article. Over time and between editing rounds on Sabati, I finally broke down and shelled for the $100/month plan. It just made more sense given how stable and reliable the output has been when I leverage the tool. Opus 4.5 has simply been amazing. It’s my go to for complex tasks in Cursor, especially with reasoning turned on. In terms of speed and accuracy it blows all the other models out of the water for me, no questions asked. Before upgrading to the more generous plan and knowing my previous API limit, I used it as that one tool you know can do real damage in video games that you consciously keep for the hairy situations before pulling out, you know what I am talking about, the Berzekator X20, with only 2 bullets. That’s Claude for me right now, I plan to use it as the daily driver for my side projects.
The team at Anthropic is amazing in my eyes, Claude evolves fast, they listen to feedback and like Cursor, there is also that element of love here for developers and building a tool that solves their problem. MCP, Skills, sub agents, they’ve been defining the game and I believe they will continue for a while.
Codex
I was initially surprised myself that Codex is coming in second because 1) it is a terminal based agent (like Claude), and 2) given everybody has been raving about Claude Code and I was not expecting to like it as much but here we are. Let me be clear, the number one reason Codex is ranking so high with me right now is two fold:
5.2 Codex
Cost
I initially did more of my terminal based coding with Claude Code, using it to learn and try to understand what the hype was all about (especially since the release of Opus). Slowly and almost in parallel, I’ve increased my usage of Codex over the past few months because I noticed the model had been getting better at coding tasks and (very importantly), on the Pro plan, Codex offers a very, very generous weekly limit. Being on Claude’s Pro plan, after 30 minutes or less of intense coding, I was pretty much at limit and had to wait a couple of hours to reset, when I didn’t bump into the weekly rate limit. If the task is important enough I would switch to my API key based auth to complete the current task, then move on to Codex to complete the task. That’s where I learned to appreciate Codex.
Codex is noticeably slower than Claude but over time I have learned to recognize it as a difference in temperament that does not affect the quality of the output ultimately. In my perception, Codex is more like a slow, deliberate and calculated sharpshooter that you can rely on, where Claude is the flashy, in your face, impressive young buck with the tricks you’ve never seen before. Today I like to use them to check each other’s work. Plan with Claude, review with Codex, and vice versa. It allows for a good measure of sanity checking when you’re working on complex features.
Conclusion
I’ll stop here and continue with the next set of tools in my AI coding artillery wall in part 2 where I’ll focus on Windsurf, Trae and Antigravity. You might ask yourself what do I do that I get to to try all of these tools deep enough to make a retour d’experience? Well outside of work, where there is quite a lot going already, AI agents have helped a former IC like me that got too busy with life, kids and management to get back to something I really enjoyed in my younger days, side projects, like Sabati i mentioned above. That’s where my man cave monthly budget goes and I get to dust off and act on all of these ideas I had now that I can execute on nights and weekends. I like the experience of trying different tools for the same tasks and evaluating how they perform, evolve over time but most importantly, enable me. After all, this is why we have this article. These projects give me enough of a variety of technical challenges that I get to spread the work out among different AI enabled IDEs and evaluate how they perform against the tasks, and how they fit in my preferred workflows. With that said, I am reaching increasingly diminishing returns in terms of productivity from being so spread out. Additionally, given the cost (and it is starting to add up), I think it is nearing time to consolidate across all of these coding tools I use to pay for the one (or two) that I derive the most value from, and go deep in configuring them to better supporting my workflows. Of course I will be sharing that along with you all.
What about you? what is it about any of these tools that make it sticky for you, what makes you return to it as opposed to any other to get your work done?