GPT‑5 Pushes Back

GPT‑5 came out last week. I’ve found it useful—but not for the usual reasons. Most AI tool launches wow you with better code or longer context. For me, GPT‑5’s upgrade is that it doesn’t over‑agree. LLMs love to help; ask them to find a bug and they’ll often invent one. I’ve chased ghosts because a model sounded confident.

GPT‑5 is the first in a while that feels different. It says “I don’t see a bug” until you show logs, a repro, or a failing test. It marks guesses, asks for missing context, and requests artifacts (diffs, traces, inputs/outputs) before proposing changes.

This week I had to reproduce two wallet UI bugs, and GPT‑5 really helped me triage and troubleshoot them. I’m sure Opus would have helped too, but GPT‑5 went a notch above and didn’t just repeat boilerplate or what I already knew.

How I use it

I’ve also started using Cursor’s chat mode again because of GPT‑5—and it’s the first time I’ve actually enjoyed the chat feature.