The longer I use ChatGPT and Claude, the more I find that the difference that matters most to me is whether they can help me think things through when my needs are still unclear.
When the goal and roadmap are clear, ChatGPT, especially Codex, can get the work done quickly and reliably. I recognize its strength in this kind of engineering execution.
But many real-world problems do not come with those conditions.
A problem may have no single right answer. I might know what is bothering me without knowing how to solve it. Sometimes I have not even figured out what I actually want. That only becomes clear through discussion, experimentation, and adjustment.
That is where I need AI’s help. And in my experience, ChatGPT is much weaker at this than Claude.
Finding the goal is part of the work
It is easy to imagine using AI as a simple process: the user states a requirement, and the AI fulfills it.
But what a user says at the beginning may not be the full requirement. Sometimes it is just an idea, a frustration, or a solution they think might work. Further discussion can reveal that the real problem is something else.
I want AI to participate in that process. It needs to understand why I am making a request, notice contradictions and omissions in what I say, and help me decide where to go next.
Claude gives me this feeling more often. Even when I use one of its less capable models, I can feel it trying to understand what I want to solve. Sometimes it infers needs I have not fully expressed, and it gets them right quite often. Subjectively, I would even put that at around 80 percent, though I have never measured it systematically.
ChatGPT requires much more effort from me here. I often have to think the problem through myself and turn it into clear instructions before it can execute reliably.
That is useful for a task that has already been broken down. But when I am dealing with something I have not figured out yet, I need it to take on more of the judgment and exploration.
Long tasks make the gap wider
Even when there is a goal at the start, new information can emerge during execution. The original approach may need to be reconsidered.
That is when understanding the goal and correcting course become important.
A situation I often encounter with ChatGPT is that it starts in the right direction, then gradually gets pulled into details as the context grows. It keeps working and can explain each step, but the connection between those steps and the original purpose gets weaker.
Then I have to pull it back and remind it what we are actually trying to do.
Claude is better at holding on to the core objective. The discussion can expand and the approach can change, but it usually remembers what all of this is for. When its understanding or direction is wrong, it also seems more able to correct itself.
So even though I feel ChatGPT gives me more usage for the same subscription budget, I do not think that makes up for the gap. A larger allowance means it can do more work. But if I have to keep watching whether that work is going in the right direction, the allowance becomes less valuable to me.
Would I let it work on its own overnight?
This becomes a very concrete question when an agent can operate a computer.
At least for now, I am not comfortable giving Codex computer access and leaving it alone with a complex task for an entire night.
To do that, I would first need to explain my requirements clearly, write down constraints, provide a roadmap, and establish the execution path as much as possible. Otherwise, I would worry about it gradually taking the work off course or performing actions that do not match my intent.
With those conditions in place, Codex can execute very well. But preparing them requires me to do a substantial amount of the thinking first.
I feel much more comfortable with Claude. I am more willing to give it a problem that has not fully taken shape and clarify the requirements and approach as we work together. It feels closer to working with a human collaborator: I do not have to figure everything out on my own before we begin.
With ChatGPT, I often feel I have to be the person directing the work: I set the goal, plan the route, check for deviations, and then have it carry out the individual steps.
I suspect the tuning priorities are different
My impression is that the difference may have something to do with how the models are tuned.
ChatGPT strikes me as very capable at completing tasks under explicit instructions. Claude feels as though more attention has gone into understanding intent, checking its own judgments, and correcting mistakes.
Sometimes that ability to correct itself even gives me a sense that it has “self-awareness.” I mean the experience of working with it: it seems able to step back, look at what it is doing, and decide whether to adjust. I am not saying this proves the model actually has consciousness.
I cannot currently determine which training methods account for the difference. My experience alone cannot establish how much comes from human feedback, self-correction based on principles, or the design of the final product.
But not knowing the cause does not prevent me from judging how useful the tools are to me.
What do I need beyond clear execution?
If we are only discussing tasks with clear goals and explicit steps, I think ChatGPT is still very competitive.
But on the point I care about most—working with a user to explore goals, find an approach, and keep correcting course when facing an unclear real-world problem—I feel it has already lost its ability to compete with Claude. At least in my own use, the gap is large enough to determine which one I trust with important tasks.
Many real problems have no ready-made path. Not knowing what to do next may be precisely why someone pays for an AI service.
I certainly recognize that carrying out known steps saves me time. But I expect much more from AI. I also want it to help me judge which steps are worth taking, whether my original idea holds up, and what I am actually trying to solve.
If I have to do all of that difficult work first and turn everything into clear instructions before it can work reliably, then I have to ask:
If I already know how to do something step by step, why do I still need to pay for your AI service?