For two years, ChatGPT was the only AI I used. After spending the last month working with Claude, Grok, and Gemini, I’m starting to question whether sticking with one AI for so long was a mistake.
I Was Getting Fed Up With ChatGPT
I’ve been using ChatGPT extensively for about two years. It’s become part of nearly everything I do with websites, publishing, software development, and technical projects. I’ve built entire workflows around it, and I’ve accomplished things I probably wouldn’t have attempted without it. That’s why the frustration I’ve been feeling lately has been so aggravating.
ChatGPT has become painfully slow at times. I’ve been dealing with too many errors, too much time correcting work that should have been done right the first time, and service outages that interrupt whatever I’m trying to accomplish. Sometimes I feel like I’m spending more time dealing with ChatGPT than working on the project I opened it to help me finish.
I expect mistakes when I’m directing complicated work. I don’t expect to hand an AI an assignment and blindly accept whatever comes back. I review results, test implementations, and send work back when it doesn’t meet my requirements. That’s part of the process. Having to repeatedly correct avoidable mistakes while waiting through slow executions is another matter.
In July, I wrote Why ChatGPT Keeps Getting Better for Me. At the time, that was exactly what I was experiencing. Each release seemed to give me more capability, and I was finding ways to use it to accomplish increasingly complicated work. I still believe ChatGPT has made enormous progress over the last two years. But more capability doesn’t help nearly as much when getting through the work becomes an exercise in patience.
More recently, I wrote about how choosing an AI model has become its own damn project. I was already spending a ridiculous amount of time deciding which model and reasoning level to use before starting an assignment. Add slow responses and repeated corrections to that, and the whole experience becomes exhausting.
I finally decided to see what the competition was actually offering. Not what the marketing said, and not what somebody else’s benchmark claimed. I wanted to find out how these other services worked with the kind of things I do every day.
I Handed Claude My Legend of the Green Dragon Project
I didn’t want to test Claude with random questions or a few simple coding problems. I wanted to give it real development work, something complicated enough that I could tell whether the difference mattered. So I moved my Legend of the Green Dragon modernization project to Claude.
I’m modernizing an old PHP browser game while trying to preserve the gameplay and behavior that made it worth saving. That involves legacy code, modern PHP compatibility, database changes, security, testing, and constant decisions about what should change and what must remain untouched. I’ve written about that work in Builder Receipt: Resurrecting Legend of the Green Dragon. This wasn’t a throwaway experiment.
I started working with Claude Opus 5.5, and I have to say, I’ve been impressed. Not because it introduced some revolutionary development technique or did something magical that ChatGPT couldn’t do. It was faster than ChatGPT Astra, and it made fewer errors. Those two differences changed the experience more than I expected.
When ChatGPT makes a mistake during a complicated development assignment, the mistake doesn’t end with one bad response. I have to figure out what went wrong, identify what needs correcting, prepare the next instruction, wait for another execution, and verify the result. If that correction introduces another problem, the cycle starts again. I’ve spent an aggravating amount of time doing exactly that.
Claude hasn’t eliminated the need for supervision, and I’m not claiming it never makes mistakes. But during this project, I’ve been spending considerably less time dealing with corrections. The work has been moving faster, and I’ve been able to stay focused on the game instead of constantly fighting with the tool helping me build it.
I now have a release candidate ready to begin testing. That’s progress, but it isn’t proof the modernization is finished. I still have to install it, run the game, test the actual gameplay, and see what breaks. A release candidate isn’t a production-ready application just because an AI says it is.
This wasn’t a controlled comparison with identical assignments, environments, and measurements. I moved a real project from one service to another, and I’m judging the experience I actually had. In that experience, Claude Opus 5.5 has been faster and has required fewer corrections than ChatGPT Astra.
For AI-assisted development the way I do it, that matters a hell of a lot more than an impressive model description.
Grok Was Interesting, but It Wasn’t for Me
I used Grok differently. I didn’t test it with code or hand it a software project. I mainly experimented with its bots to see whether that kind of assistant would be useful in my day-to-day workflow.
I’ve also spent time with ChatGPT’s Dots, and I actually liked the Grok bot experience more. But liking the experience and needing the feature are two different things. I already have ways of planning, directing, reviewing, and maintaining my projects. With the way I work, bots don’t add much value to that process.
I’m not opposed to tools that work in the background or help keep track of ongoing work. They just need to do something useful enough to justify making them part of my system. Otherwise, they’re another feature to configure, supervise, and think about without much to show for it.
There was also something about the overall Grok experience that didn’t sit right with me. I can’t point to one specific failure that explains it, and I’m not going to invent one. I simply didn’t like using it enough to continue.
Because I never tested Grok with actual development work, I have no basis for comparing its coding performance with ChatGPT or Claude. Someone else might get a lot out of it. I’m not going to use it.
Gemini Made GitHub a Pain in the Ass
I wanted to test Gemini with code, particularly projects I already maintain through GitHub. That should have been a reasonable place to start, considering how much of my development process revolves around repositories, branches, requirements, and testing.
One of the things I appreciate about ChatGPT and Claude is being able to work with GitHub as part of the development process. I can direct work against an existing repository, review changes, and keep the project moving without manually shuffling code around. I’m not looking for an AI that just spits out random snippets. I’m working on applications that need to survive beyond the conversation.
With Gemini, I couldn’t get that same kind of integrated GitHub workflow working the way I needed it to. I had a hell of a time figuring out how to connect the pieces, and it became more frustrating than the coding work I wanted to do in the first place.
Maybe another Google tool or configuration would have solved it. I’m not claiming Gemini can’t work with GitHub in any form. What I’m saying is that I couldn’t get a workflow comparable to what I already had with ChatGPT and Claude, and I wasn’t interested in rebuilding how I work just to accommodate another service.
Eventually, I gave up. That meant I never got far enough to make a meaningful judgment about Gemini’s coding abilities. The workflow stopped me before the actual test could begin.
I did keep using Gemini for some research, and that experience has been pretty solid. I can see a place for it when I’m looking into a subject or working through information. But for my development work, a GitHub connection that fits how I actually operate isn’t an optional convenience. It’s part of the job.
ChatGPT Still Has a Place in My Work
I haven’t abandoned ChatGPT, and I don’t intend to pretend that the last two years were wasted just because Claude impressed me. ChatGPT has helped me build websites, develop software, work through complicated technical problems, and establish publishing workflows I still use.
It knows a great deal about my projects and how I work. I’ve invested time in establishing that context, and I’ve built systems around it. I still use ChatGPT for WordPress, the Proof theme on JimLunsford.com, publishing, book projects, and server administration. For much of that work, it does a decent job as long as I don’t expect quick results.
Those capabilities are real, which is why this isn’t some dramatic announcement that I’m deleting my account and starting over. ChatGPT remains useful to me. I’m just a lot less willing to accept the same frustrations when I’ve seen another service handle development work more efficiently.
There’s also the matter of what I’ve already invested. Two years of project history, established workflows, and familiarity with the tools don’t transfer overnight. Starting over somewhere else would be a substantial job in itself. That’s probably one of the biggest reasons I haven’t considered switching completely.
But I have to be careful with that reasoning. Staying with something solely because I’ve already put time into it isn’t a good reason to keep accepting problems. If I were starting from scratch today, knowing what I know now, I’m not sure I would put ChatGPT at the center of everything the way I did two years ago.
I don’t need to throw out the parts that work. I do need to stop assuming ChatGPT should handle every new project simply because it’s what I’m used to.
I Don’t Need One AI to Do Everything
For two years, my default was to open ChatGPT whether I was writing an article, building software, managing a website, working on the VPS, or researching something. I didn’t have much reason to question that habit because it kept helping me do things I couldn’t have done as easily on my own.
The last month changed that. I’m less interested in which company claims to have the smartest model and more interested in which tool helps me accomplish the work without introducing unnecessary friction.
Speed matters when I’m waiting on a complicated assignment. Error rate matters when every correction requires another round of instructions and verification. Integration matters when the entire project already lives in GitHub. None of those concerns disappears just because a model looks impressive on paper.
There’s a difference between complexity that belongs to the project and complexity introduced by the tool that’s supposed to help me. I know that modernizing a legacy PHP game is complicated. I expect to have to think through architecture, testing, and security. I don’t want the AI itself making that work harder than it needs to be.
So my current division is pretty simple. Claude is my preferred choice for serious AI-assisted development because Opus 5.5 has been faster and more reliable in the work I’ve given it. ChatGPT still handles my established publishing, website, book, and server workflows. Gemini has been useful for research, while its GitHub experience kept me from properly testing it for development. Grok doesn’t fit how I work.
Those are my results, not a universal ranking. I haven’t tested every model, feature, or workload. I also don’t expect today’s preferences to be permanent. If one of these services improves or another one gives me better results, I’ll change what I use.
I’m not looking for a single AI company to win some imaginary loyalty contest. I’m looking for tools that let me spend more time building and less time dealing with the tools.
ChatGPT Has Some Catching Up to Do
A month ago, I wouldn’t have expected to say Claude was giving me a better development experience than ChatGPT. Now I am. For the kind of AI-assisted software work I do, ChatGPT Astra has fallen behind Claude Opus 5.5 in speed and in the number of errors I’ve had to correct.
That’s disappointing after two years of relying so heavily on ChatGPT. I still appreciate what it has allowed me to accomplish, and I still see a place for it in my work. But appreciation doesn’t require loyalty.
If Claude keeps giving me better results for development, I’ll keep using Claude. If ChatGPT becomes faster and more reliable, I’ll recognize that too. I have no interest in defending a company or convincing myself that the tool I’ve used the longest must be the best.
I’m interested in getting the work done.
After two years of relying almost exclusively on ChatGPT, branching out was overdue.
New Here?
Read Next:
- Why ChatGPT Keeps Getting Better for Me
- Choosing an AI Model Has Become Its Own Damn Project
- How I Actually Build Software With AI