anyone else used gpt-5 high as a daily heavy lifter?
I find it to actually be the best model im working with now. in terms of cost its affordable too and its practically the best when it comes to following instructions and also the best when it comes to using the edit tool, barely any mistakes being made. Just wish it had like double the context that it has now.
It’s a main model I use along with codex. Slow but incredibly smart! And works great with new Plan mode in Cursor
4.5 sonnet
is the biggest letdown of the year. gave it a task and it couldn’t even use the edit tool correctly, always made a mistake and just overall got confused. it was so depressing i started thinking they lied with their evals cause gpt 5 high
is “dumber“ than it on evals but has never failed to complete a task correctly.
gemini 2.5 pro
is only if i have too much context, and its kinda 50/50 on following instructions and deducing errors. sometimes it works so well you really do start seeing why its ranked as the best reasoning model out there. then most times it just flat out hallucinates lol.
overall your deduction is spot on. also why did you downgrade from the latest cursor build?
I love Gemini, but it’s really outdated and hasn’t been working very well as an Agent since its release. I’m really looking forward to Gemini 3.0 Pro and I am very afraid that it may work poorly in Cursor again, or even not work at all.
lets just hope its more of what gpt 5
was and not claude 4.5 sonnet
. we’ll just have to wait and see, but the frequent cursor updates arent helping with fixing these issues
its honestly not just you. 1.7 is buggy as hell. experienced a couple bugs myself
I can tell you that it fixes bugs better than Claude 4.5 or any other AI, and sometimes even pinpoints single-line problems.
also sonnet is much more expensive than gpt 4 high
I’m not a big fan of claude sonnet 4.5, but still using 4.
Im testing Codex atm and asked the AI to “fork” colors and setup from the lambo. landing page.
Result:
- Planig: seems strong
- Finale result: I had to do 5 requests until it was error free. It seems that codex is still not that good for code writing for me.
Cannot fix such import traces etc-- with this reason I tried GPT-5- High Thinking model.:
- Planing seems logical
- Final result: 2 requests and still not operating, im switching to Sonnet 4.
Sonnet 4:
- Planing: analyzed the errors and made a plan how to fix it
- Final result: one shot and landing page is up and running
So for me personally, sonnet 4 is still the “king” under the llms for coding (speacially with react)
Maybe I’m missing the part “how and why” we use GPT-5 high or codex, but for me I will keep with Sonnet.
Edit: But the design edit by Codex & High is impressive, better than sonnet
what about high fast ? Does it give the same result faster or worse?
its same with high but price is 2x because of speed