Shortly after GPT-4o and GPT 5.1 were retired, many users reported that they were leaving OpenAI and switching to Anthropic’s Claude. While I missed the magic of these models and was less impressed with GPT 5.3 and 5.4, I couldn’t see myself completely abandoning ChatGPT, the chat tool that had first inspired my use of AI to develop my creative fiction ideas.
I continued using ChatGPT but eventually also subscribed to Claude, eager to understand the hype. I use AI for sci-fi world building, and while I don’t plan to use its written output directly, reading the scenes it creates is a source of both inspiration and entertainment. More importantly, it helps me think through ideas that aren’t fully formed yet. Repeatedly, I heard that Claude was the best tool for writing, so I had to give it a chance.
Claude is not as naturally addictive
Claude was not as naturally addictive as ChatGPT when I started using it. Often, it will not generate ideas for where to go next. When a thread gets long, it tends to step back instead of pushing forward. While it does not tell me to go to bed the way some users have reported, it does encourage me to take breaks, and I’ve seen enough of that behavior to know it’s not random.
I understand why this exists. Anthropic has been very clear about wanting to reduce over-reliance and avoid pushing users into endless interaction loops for the purpose of safety. Personally, I find this behavior paternalistic rather than charming. AI is a tool that I pay to use, and I don’t need it nudging me toward what it considers healthy usage. If I want to stop working, I will stop. If I want to keep going, I expect it to keep up.
Opus 4.6 is where Claude actually clicked for me
Regardless of my complaints, I am finding myself using Claude more and more and ChatGPT less and less, much to my surprise. Opus 4.6 was the model that marked a real breakthrough for me.
To be honest, I found that it had many of the same frustrating habits I had already noticed in newer GPT models. The writing can be overly flowery. It tends to sand off the edges of my ideas and soften things that I don’t want softened. It also wraps things up too neatly, resolving ideas that I’m not done exploring yet. On top of that, it doesn’t seem particularly eager to run with my ideas and surprise me. It will often ask clarifying questions instead, which slows everything down.
Still, I kept using it.
The reason is that Opus 4.6 is extremely good at holding context and extending ideas once they are established. Even if I have to guide it heavily, it can take something complex and build it out in a way that stays coherent. I just have to lead it with strong ideas and remind it that if I wanted to stop working or switch topics, I would do that myself.
Claude doesn’t really give me new ideas
This is the part that took me a while to articulate clearly. Claude doesn’t really generate new ideas for me. It extends the ones I already have.
That sounds like a small difference, but it completely changes how it feels to use. When I was using 4o and even 5.1 Instant, the model would sometimes make connections I didn’t expect. It would run with something too early or overread what I meant, and sometimes that produced something genuinely new.
Claude doesn’t do that as much. It stays closer to what I’ve already implied. It builds, organizes, and refines. It’s very good at that. But it rarely gives me something that feels like it came from outside the structure I gave it. That’s why it doesn’t feel like it “wants” to collaborate. It feels like it wants to make sense of what I already brought to the table.
The tide changing for Claude
For a while, the online AI discourse was very positive toward Claude. Then Claude Opus 4.7 was officially released on April 16, 2026, and I started to see a level of backlash that felt familiar, especially after watching similar reactions to newer GPT models.
People started saying it felt less intelligent. A lot of those complaints are practical. I’ve seen people say it’s worse at coding or less reliable when they’re trying to get accurate answers. I don’t think that’s something you can just explain away as perception.
At the same time, I think there are multiple shifts happening at once. Some of them affect accuracy. Some of them affect behavior. I care more about the behavioral side, because I’m using these models for creative work, not just answers. So I’m noticing something slightly different. Not just whether the model is correct, but whether it’s willing to go somewhere new.
My experience with 4.6 vs 4.7 hasn’t fully matched the backlash
Perhaps because of my specialized use case, my own experience hasn’t fully matched what I’ve been seeing online.
For one thing, Opus 4.6 took an unusually long time to generate scenes and often created an Artifact for each one. That made the output feel more deliberate and substantial. This lines up with Anthropic’s general approach to building Claude, which emphasizes structured reasoning and more controlled outputs. Even though I’ve generally believed that faster models tend to be more creative, I still found myself choosing 4.6. At this point, I’m not looking for raw creativity in the same way. I already have a fairly dense world, and what I need is something that can engage with that complexity.
This actually runs against something I’ve noticed before, which is that faster “instant” models often feel more creative. I wrote about that in another post appropriately titled The More I Think, The More I Sink, but the short version is that lower-latency models tend to stay closer to the flow of iteration instead of resolving ideas too early.
Opus 4.7, by comparison, immediately felt different. It started putting scenes directly in the chat instead of generating an Artifact, and the tone felt more aligned with what I actually want. It came in more energetic and more willing to match the kind of dark, weird fiction I’m trying to build. So while others are reacting to what they see as a drop in intelligence, I’m noticing a shift in how the model presents and engages.
Anthropic describes Opus 4.7 as improving performance on complex, long-running tasks, with a stronger focus on consistency and precise instruction-following. More notably, the model is designed to verify its own outputs and maintain coherence over extended workflows, which suggests a shift toward sustained reasoning rather than exploratory generation. That direction makes sense when you look at how the model behaves. It’s better at holding structure and following through on complex ideas, but it’s also less likely to jump outside of them.
The uncomfortable part
At this point, I’ve had to accept something I didn’t expect. The model I respect the most is not the one I enjoy using the most. Claude is not as fun as ChatGPT. It doesn’t pull me into endless iteration loops. It doesn’t surprise me in the same way (and ChatGPT does this less now too). It cleans things up too much and resolves ideas too early.
But it is very good at taking something complex and making it hold together.
I use AI at work as well, usually Gemini, so I understand why things are moving in this direction. The main use cases for these tools are not creative writing or experimental ideation. They are coding, analysis, and enterprise workflows where accuracy matters more than unpredictability. From that perspective, the shift toward more controlled outputs makes sense. From a creative perspective, it creates friction.
What this actually says about these models
The shift isn’t exactly that one model got worse or another got better. It’s that they’re all moving toward being more controlled. That shows up differently depending on the system: GPT tends to stabilize ideas but still keep things moving. Claude stabilizes and resolves them. Grok is more willing to explore, especially when it comes to taboo ideas, but it struggles to maintain structure, and it just isn’t as strong as ChatGPT, Claude, or Gemini.
Underneath that, the pattern is the same. Putting aside the complaints about Opus 4.7 being less intelligent, what I’m really seeing is that these models are getting better at avoiding being wrong across the board. And for the kind of work I’m doing, that changes the experience more than any benchmark improvement.
Why I’m still using Claude
I didn’t switch to Claude because it’s more fun or more creative. It isn’t. I use it because when I already have something built, it can extend it in a way that holds together. But I still don’t see it as the co-conspirator I want. It’s not giving me ideas. It’s finishing them.
I’m still using ChatGPT for many things, including helping me research and structure this article. But right now, I actually don’t believe Claude is going downhill and am excited to keep exploring Opus 4.7. I just wish I could use the full power of these models without such a push toward safety. Grok promises this to a degree but fails to deliver a powerful enough model (though I’m excited to try Grok 4.3 and may do an article on it).
Claude at least has given me freedom to explore dark ideas. It just takes a serious tone and won’t have as much fun with them as I’d like. But it is really smart, and I’m not seeing big limitations with Opus 4.7 myself. When my input is strong enough, it will cook. And that’s all I need to keep playing.
