@maxast.eth
I've been on the Codex Pro Plan for the past 1-2 months and I'm extremely unhappy with the code quality lately.
I went for it because I kept hearing that GPT 5.3 is the smartest model, but it's sooo bad at following existing code patterns and style. I use Effect heavily and my codebase is pretty clean (if I may say so), so in theory it should be really easy for a frontier model to just follow the patterns when adding new endpoints/services/db queries etc.
In fact, it has produced so much slop that it has probably been a net negative on my productivity over the last 2-3 weeks.
What's funny is that I started letting Opus 4.6 review GPT-5.4's code yesterday and it sees the exact flaws that I see - but when I let a fresh GPT-5.4 session review the code, it just doesn't see the flaws.
GPT-5.(3|4) is great for finding logic flaws and bugs, but it's really bad at being a productive software engineer imo.
0 reply
0 recast
0 reaction