I know GPT is the favorite model family for a lot of people. I’ve tried to get there. I really have.
I’m running GPT-5.6 Sol right now, and it’s still the same experience I’ve had with every GPT version I’ve touched. The model runs forever on simple issues. There’s hardly any reasoning output, so you can’t see whether it’s thinking in the right direction. And when it finishes, you get three words:
Slice done.
That’s it. No context about what changed. No mention of what didn’t work and why. No instructions on how to test the result. Just a binary signal: task complete, trust me.
The Black Box Problem
I don’t need my models to be chatty. I’m not asking for a paragraph of self-congratulation after every file edit. But there’s a world of difference between concise and opaque.
Compare that to GLM-5.2, which is my daily driver right now. When GLM finishes a task, it tells me:
- What it actually did
- What it didn’t do and why
- How to verify the result works
That’s not verbosity. That’s auditability. That’s enough surface area for me to judge whether the work is sound before I commit to it.
“Slice done” gives me nothing. It’s a black box, and the only way to know if the output is any good is to open every file and check by hand—which defeats the purpose of having an agent do the work in the first place.
Running in Circles
The other thing. GPT seems to loop.
Not agentic looping—that’s a different problem. I mean the model appears to run in circles during a single task. What should take thirty seconds takes two minutes. You can see the token counter climbing and climbing, and you start wondering: is it thinking deeply, or is it stuck? With GLM, I can see the reasoning output. I can watch the model work through the problem in real time and catch it early if it’s going down a wrong path.
With GPT, you’re just… waiting. Hoping. And then “slice done.”
Maybe the output is perfect. Maybe it’s garbage. You won’t know until you look, and the model has given you no reason to feel confident either way.
Trust Is Expensive
This isn’t a benchmark. I don’t have logs, side-by-side comparisons, or carefully controlled test cases. This is just what it feels like to use these tools day after day.
And what it feels like is this: I trust GLM because it shows its work. I don’t trust GPT because it doesn’t.
That might seem like a soft reason to prefer one model over another. We’re supposed to evaluate on benchmark scores and reasoning capability measured in carefully designed tests. All of that matters, I’m sure. But none of it matters as much as the daily experience of sitting down to work and reaching for the tool that doesn’t make you second-guess every result.
Communicativeness isn’t polish. It’s the difference between a tool that works for you and a tool that makes you work to verify it did anything at all.
GPT might be brilliant under the hood. I can’t tell. And that’s the problem.