Earlier this year I wrote a few posts about switching from ChatGPT to Claude, then getting into MCP, and various odds and ends I noticed after making the move.
Quick background, in case you're new here: I'd used ChatGPT as my main AI tool for a long time, but for various reasons I switched my main tool over to Claude around spring 2026. (I haven't quit ChatGPT, by the way — I still use it as a secondary tool.)
It's already been a few months since I moved over to Claude. (Time flies 😅)
I happened to be looking back over one of my older posts, and now that I've got a few more months of use under my belt, some new things have come into view — stuff I couldn't have noticed back when I wrote that earlier post. So this time I want to touch on that a bit. This isn't so much a tidy conclusion as just my scattered impressions right now.
I previously wrote that the combination of the memory feature and MCP "cut down on explanation overhead quite a bit." That assessment still holds. But the more I use it, the more I've started to see another side. For example, sometimes it'll force a connection to information stored in memory even when the topic is completely unrelated.
"Uh, that's... not relevant right now."
Moments like that feel more frequent than before. These forced connections add noise, and it can get a little annoying. It's the flip side of the convenience, but since memory just keeps accumulating, tidying it up and deleting things might become part of the routine going forward. (I haven't tried that yet, so it's a to-do for later.)
One more thing — and I can't be sure about this — but lately I feel like I'm running into a bit more hallucination, excessive agreement, and, as a result, a harder time thinking critically. Even with Opus 4.8 — the top model if you set aside Mythos and Fable — I occasionally hit the kind of wild, off-base answers I rarely used to see. (Oddly enough, when I experimentally switch back to something like Opus 4.6, I sometimes feel the answers are actually better, which is strange.)
It might have to do with how well a given model fits the kind of thing I'm asking about. Or maybe I'm wrong and it's just my personal impression, or maybe it's something off in the model's design itself — I can't really tell. Specs change often, so it could just be a problem with how I'm using it. I'd like to keep watching for a while.
Right now I'm using ChatGPT on the free plan. Back when I was weighing the move from ChatGPT to Claude, I agonized quite a bit over whether to keep my ChatGPT Plus subscription. But paying a fixed cost for something I'm no longer using as my main tool felt wasteful, so I decided to try going free.
The number one thing on my list of "what ChatGPT has that Claude doesn't" is image generation. Claude can handle things you can express through SVG or Artifacts, but it doesn't have standalone JPG image generation the way ChatGPT does (as things currently stand). In my case, I need this image generation in all sorts of situations, so I was worried about whether going free — with tighter limits — would actually let me generate the volume my work requires. (Or, put another way, I figured if I hit a wall here, I'd need to keep Plus after all.)
But when I actually tried it, I could generate a fair amount without hitting a stop. The quality looks unchanged too. That said, if I ever need a much larger volume of images in a short window, things might change.
For now, though, the volume I use for work is no problem, so I feel like staying on the free plan is fine. On the whole, I mostly use ChatGPT for image generation these days, but it seems like new models and memory features keep getting added one after another. There are probably quite a few features I don't even know about.
As for the recently talked-about GPT-5.6, I've also seen reports that on some benchmarks it's putting up numbers that surpass Claude's Mythos class — the model that itself made waves around the world.
So my impression that "Claude is ahead" has faded compared to before. But in the end, both are evolving at a pretty rapid pace, so I suspect my impression will have shifted yet again six months from now.
Today was all over the place, but I've loosely gathered up some honest impressions of how AI has felt to use lately.
One thing that struck me is that my own workflow is changing even faster than the AI itself. Properly adopting and using MCP, for instance, dramatically changed how I run my work, and going forward I plan to keep trying out new ideas — not just MCP — as they come up.
If I read this back again in a few months, I might think, "Huh, so this is the stuff I was fretting about back then." This post, too, isn't a conclusion — I'm leaving it as a note capturing where things stand right now. Thanks for reading!