When GPT-5 Becomes That Talented but Annoying Teammate
My love-hate relationship with GPT-5 in Cursor — gorgeous code, but a habit of making things that don’t actually work unless I crank it to MAX mode.
Look.
I wanted to love GPT-5 in Cursor.
And at first glance, I did. The code? Beautiful. Like, “I’d hang this on a wall if it were socially acceptable” beautiful. Indentation perfect, naming conventions on point, just the right amount of elegance in the logic. It’s the type of code that makes you think, Damn, this AI gets me.
But here’s the problem: it doesn’t always work.
Unless I crank it up to GPT-5-high MAX mode — which is overkill for small tasks — the thing has this infuriating habit of making stuff that looks brilliant but falls apart faster than a cheap knockoff. And MAX mode isn’t exactly something I can use for everything. Some of my requests are small, tactical. I don’t need to unleash the AI equivalent of a nuke just to fix a conditional statement.
Worse? Half the time, Auto mode actually does better.
Which makes me wonder if GPT-5 is like that genius teammate who insists on rewriting your entire module when all you asked was to change a single line. It’s not laziness — it’s performative genius. And I’m the poor soul left refactoring the aftermath.
And here’s the kicker: I end up having to pay for Sonnet 4 just to fix GPT-5’s work. Think about that. I’m using one premium AI to clean up after another premium AI. If that isn’t peak 2025 developer absurdity, I don’t know what is.
Now, here’s the thing — I gave GPT-5 all the chances in the world.
I was hyped. I used it for everything I did. I ran my usual workflows. I wanted it to be my new daily driver. But when it came to actually producing working code, it lacked that final oomph.
Where GPT-5 absolutely shines?
- Boilerplating code
- Documentation
- Understanding a codebase in seconds
- Writing clean markdown files
- Mermaid.js diagrams on demand
For that stuff, it’s amazing. I’d still happily use GPT-5 as my documentation powerhouse. But for raw coding performance? Others just do it better.
The problem is… once Cursor starts charging for GPT-5, I can see the hype dying fast.
It’s hard to justify paying for something that writes beautiful but unreliable code, especially when you’d need a GitHub Copilot subscription to get the most out of it anyway. (Which I do have, by the way.) At that point? I might as well just meet GPT-5 in VS Code instead.
So yeah — until it can match beauty with reliability without MAX mode, it’s staying in my “talented but annoying” category. And the day Cursor puts a paywall on it? That’s the day I quietly close the tab and never look back.
~ FIN ~