Your thoughts on GPT-Codex-5.3 vs GPT-5.4 for engineering/coding tasks..
#What do you prefer for coding/engineering GPT-Codex-5.3 or GPT-5.4 and why?
jagged tree
fast forge
5.4 xhigh
royal delta
so far I prefer 5.4, it feels like it reasons more about an issue and looks more at the bigger picture compared to 5.3 codex. I'm trying to get it to a complicated task in OMSI 2 where it adds an umbrella model to humans roaming the map in rain or snow. This involves using memory manipulation and hooking directx. Undecided between xhigh and high currently. Voratiq's benchmark claims they are both very close with high slightly in the lead right now. Currently using xhigh
also fancy seeing us three here (from the other discord 😂)
royal delta
ah maybe not, YOI I know in another discord then, maybe he's not in the one I'm referring to
jagged tree
so far I prefer 5.4, it feels like it reasons more about an issue and looks more...
interesting, and which harness do you use? codex app or cli ?
royal delta
interesting, and which harness do you use? codex app or cli ?
right now I'm using opencode as I don't trust using Codex temporarily with the recent usage consumption issues reported
fast forge
opencode not taking up much usage?
royal delta
opencode because it has DCP plugin and I can use a hook to use rtk, which both ideally save tokens
royal delta
opencode not taking up much usage?
nope seems pretty slow, I'm still on 100% despite using it for an hour this morning 😂 (weekly)
jagged tree
and do you really notice a difference in token usage with the DCP plugin even with the 2x OpenAI offers?
royal delta
while it's a perception, a feeling, not actual concrete data comparison, when using DCP it has felt like my token consumption is slower
less auto compactions of course
both now and past usage that is
fast forge
nope seems pretty slow, I'm still on 100% despite using it for an hour this morn...
they must have fixed it
royal delta
yeah they did do some changes and those who shared their user ID in that GitHub thread got a nice reset too 🙂
royal delta
a bit late but maybe there's another chance of a reset next week 😄
royal delta
I see
fast forge
70% remaining I mean
royal delta
yeah I know lol
celest adder
I tested GPT-5.4 (High) alongside with GPT-5.3-Codex (High) on a non-toy project for several ours with many samplings over the same prompt and different issues.
I assumed that GPT-5.4 might perform better, but I'm a little bit disappointed.
In most cases GPT-5.3-Codex wins.
My conclusion is: GPT-5.3 explores the code base much better and comprehend the data and execution flow for the core problems and detects also some inter-dependencies better, whereas GPT-5.4 misses parts of the code base during discovery.
Codebase exploration is a critical part to be familiar with the slices at hand and finding issues, bugs and implementing features.
I will stick with GPT-5.3-Codex until real tests showing better results and the hype decreases with real-world sample results.
And GPT-5.4 is too much verbose, it concludes with a wall of text I don't want to read. Where it really shines is by generating documents and design plans. But here I doubt that they are correct, because of missing essential parts in the codebase.
jagged tree
I tested GPT-5.4 (High) alongside with GPT-5.3-Codex (High) on a non-toy project...
I see, and which programming language(s) did you use it for ?
celest adder
In this code base the following
