this post was submitted on 20 Feb 2026
2 points (75.0% liked)
Lobste.rs
330 readers
29 users here now
RSS Feed of lobste.rs
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Opus 4.6 and GPT 5.3 Codex produce some amazing results if you spend a lot of time scoping, speccing and testing each stage. Just throwing it over the wall with a low-effort prompt isn't going to get you anything that's very good. But time spent on the leadup and close monitoring of the results can give you production-ready code with little tech debt in it, extremely quickly and without a lot of money (energy) spent on inference.
Now downvote away.
Just today, I asked GPT5 mini for a little side project to give me a python function that returns the access rights to a given file/folder, using smbprotocol. For me, that read as a pretty concise ask, but the results were always using functions/attributes that didn't exist.
Any time I've asked for scripts, it's been flawless and way more than I asked for, usually with switches like --host, key, auth etc. But I'm using at least Sonnet if not Opus. I'd punch out a script for you with it but I have nothing that uses SMB in my network to test on.
If you're going to use GPT, you want 5.2-Coder at least, and honestly I'm not as impressed with OpenAI's products as other people seem to be.