GPT 5.2 is "ridiculously better" than Opus 4.5 they said... I tested BOTH on identical bugs in OpenCode and recorded everything GPT 5.2 → 24 minutes for nothing Opus 4.5 → 4 minutes and done Full video proof (crazy) Always test yourself instead of trusting tweets

You can't test this in opencode, it's not the same thing. Harness is important. 5.2-codex IN codex. Rest isn't as good.

BTW, just did the same test using Codex, and it's way faster! 6 minutes instead of 24, but it didn't fix the bugs. Using "high" mode.

For stuff like that, consider closing the loop so codex can verify this end-to-end, then it'll fix it for sure. Time-wise, Opus will always win. Codex rules because it doesn't miss stuff in large codebases.

i agree, but if you do a "loop" with result for Opus, Sonnet or even maybe Haiku it will work after some times

BTW, I did more tests. This is the FIRST test of a series of tests to see the diff with same prompts/same env. Just sharing this one because I found it CRAZY that it can spend 24 minutes for nothing. This was just unexpected

Important: I do not say that Opus is always better. I am saying that there is a balance. The Paul post shows a strong opinion: > Codex 5.2 is BETTER than Opus 4.5 Yes, maybe sometimes, but this is not a "bold truth." Just try it yourself and get your opinion. For me, speed is really important, and Codex is unusable just because of that.

full video here :

update : gpt 5.1 codex high make it ! in 12 minutes but works

Interesting - my experience was the opposite actually GPT 5.2 handled my codebase better on a similar debugging tasks Probably comes down to the type of bug, language, or context size. These models have different sweet spots Agree 100% on testing yourself though

yes, it's depend. What i like about opus is that he follow what you ask him, most of the time where gpt seems to be more following is own things

@prodhi_code Given your broken English, maybe it’s just that it’s better at interpreting broken English or whatever your native language is.

@prodhi_code Consistent disfluencies.

@prodhi_code You are a smart tech lead, I hope you support your team the same way, even with their accents.

could have a million counter proofs honestly can’t count how many times claude models (including Opus 4.5 Thinking) have messed up shit the most ridiculous behavior that is still in even the latest and greatest, is how often it will take the easy way out either reduce plans to “make it work” or doing extremely unsafe things like a login that’s tested by setting a cookie instead of any backend logic (real example) stay away from claude models if you are not a coder and ship production apps

The funny thing is that all vibe coders creating apps use Claude since the beginning, but yes, yes. And yes, we can have many counter examples; I just want to show the "bad" about this kind of tweet

a lot of people used it and still use it every day. it’s not a bad model (would never claim it is). but it’s just not as good as codex now. that was obviously different with 3.5/3.7/4.0, but since then things have changed quite a bit i use opus to a lot of things, including some minor code. but mostly for writing and agentic stuff i do every day. i think if codex genuinely pushes into cerebral, you’ll see anthropic lose out faster than the speed of light speed is still the number one thing people complain about

but i agree, never say that gpt 5.2 is, my ONLY args is this : they are as good each other

they both have good things, yes but one is better for the majority of coding tasks, especially if you want to ship real things that are not just a website or a smaller app

test in codex, dont use opencode. opencode sucks with gpt 5.2 rn

And Codex is "officially supported" in OpenCode, so should be fine

ok i have two gpt pro subs im telling you and two CC max subs but you do you!

yes I have Codex sub too, I use it in OpenCode to have the same env, same system prompts, same tools to tests the models, not the env ;)

what more can i say 😅

you tested it on one bug? your basing debunking people’s opinions on a sample rate of one

How about you test it on 100 bugs and see which one fixes it more often. Some anecdotal one-off means nothing dude

you are completly right

Thanks for providing the other side. Opus is capable, but it’s not good at following directions. It consistently broke my rules in my Claude .md file. That’s why I don’t use it. It says it does a task and it doesn’t. It gives up early or it does something I say not to do.

opus: verbose and thorough gpt: concise and sometimes wrong the question isn't which is better—it's which failure mode you prefer

i don't know if gpt is less verbose, thinking block is huge

that's why every time i say to my friends and others, "don't believe the benchmarks too. test yourself, then see if you like it or not."

i can only agree

I find GPT 5.2 to be better at debugging, but not necessarily fixing. Using a mix of both has helped me fixed a lot of issues Opus could not.

Farming engagement slop artists post the same thing every time a new model is released 'Xyz is DANGEROUSLY good' Patiently waiting for opus 4.6 is RIDICULOUSLY better than codex 5.2

Listen carefully. 5.2 codex xhigh. The think you use is everything. Gpt 5.2 on what level? Low, med, high, xhigh? Because opus to codex xhigh is like gpt4o to sonnet 3.5. No one would ever say gpt4o could code worth a damn

i try the "high" because xhigh was too slow to even finish the task 😂

I have four max accounts. And four pro accounts that I use almost every day. Opus, even with guard rails is often incomplete and inconsistent. Codex Xhigh is always cleaning up its mess and soon it will be running async parallel agents out of the box. Used Codex Xhigh in gastown last night and it was magic. Enterprise level infra at scale. 14 agents with gating and extreme gating. Freaking wonderful, emasculate code. Opus? “Fire Ready Aim”. Incomplete. Codex? ……. …… …… 20 Minutes. “Work’s Done!”

With my workflow, opus NEVER stop his tasks. Like he always get 200k context limit before giving up.

Oh I always run out of context lol. Like before this most recent update with search mcp I was constantly have my poor sub agents die in the field and my orchastrator being unable to compact. Not lazy in that sense more like not thurough enough and maybe I am scoping to big of tasks for them. I am not a coder. So I look at other things. Like one, does it work? And twos is the code structure consistent accross functions so I know based on geometry it’s ideal or minimal. But in terms of work load I don’t really think about I just try to see what the models can do. On average I forget to clear codex and it at times has had nearly 400m in cashe…. It gets real myopic at that point

They’re just Sam’s shills 😂

garbage slop larp rage bait

Melvyn, your testing approach is crucial for understanding the true performance differences between models, and this is why performance benchmarks.

😂 thanks you

Eh, I run into things each solves the other can’t. As long as it gets done and I have the means of doing it is the most important outcome for me. I think we’re headed for a world where the models are all abstracted and it doesn’t matter, but until then keep leveraging your arsenal to its fullest.

Yup, I use both in Opencode. And when Opus can’t solve something I sometimes switch to GPT 5.2. But GPT does sometimes get stuck in a loop and just end up weirdly outputting some kind of neverending haiku.

that grammar check with @raycast is beautiful - didnt know that 😂

@raycast Ahaha, yes, this is my hack to not make 100 mistakes in a video.

5.2 is meh. 5.2 high significantly outperforms opus. Actually not even close from my expeirence. Opus is fast and probably better in certain situations, it's all I used until a few weeks ago IMO. Playing with 5.2 codex high and actually seeing it underperform 5.2 high oddly.

bro look at the video, I USE 5.2 HIG 😂

Don't use it then 🙃

Here is the simple truth: If you are building at the most cutting edge with incredibly difficult problems, GPT 5.2 XHigh beats Opus almost every time but it takes 5x as long. So for many other coding problems, Opus is the better one to use (simpler tasks that don’t require the intelligence)

the "truth", you're truth yes, actually in this really difficult and tricky problems, like really difficult to see, i would take year to figure out, Opus fix it. But yes , in some case, gpt 5.2 is better

@MattBushND What Matt says is correct.

@MattBushND Are you got of truth, there is no truth, it’s opinion of people

rage bait

😂 love u too

Disinformation wars. Many people are making it ideologically motivated in a sense picking a “side” and pushing a message rather than pragmatically explaining the value they each bring.

yes exactly, didn't see any real post with example where gpt 5.2 is way superior

It’s not. It just good at getting poorly skilled less knowledgeable people engaged.

If even Elon posted on X that grok 4.2 will not win againt antrhopic on coding because anthropic has done something special with their coding abilities THEN you know athropic is king... openai will just not admit...

This is why “model X is better” debates are mostly noise. The winner is the one that solves _your_ bug class with the least retries + context wrangling — time-to-fix is the metric that matters.

GPT 5.2 is the goat.

the what ?

GOAT. A phrase where we have a phrase with four words beginning with G, O, A, T. This stands for greatest of all time. The absolute best, the absolute best at something.

@grok, how can @melvynx perform a power analysis so he can determine how many eval runs he really needs for a statistically significant outcome?

You might have conveniently tested in openCode probably with a poor environment set-up. Test it in codex cli and "record everything" 😂😂

today is you're day, i did it :

Nobody said it was faster. It's slow, but sometimes a bit more clever.

but in my case, it's slower and worst, difficult to know when to use it

Paul now show us your homework

this is exactly why I run my own tests before switching models. the hype cycle is real but the only metric that matters is: does it solve MY specific problem faster? 24 min vs 4 min is brutal lol

Speed shouldn't be your first metric. It's not about speed. If you tell a model find a cure for cancer and its taking months but comes up with correct answer you wouldn't mind

but for code i want speed. Speed is one of the most important metric, slower mean less review, less feedback everything

I share the same impression: while both GPT and Claude perform well, Claude consistently delivers results that align more closely with my expectations.

Been my experience as well

GPT 5.2 (codex) IS RIDICULOUSLY BETTER… at spending 20 minutes exploring code just to come back with “Search complete. Would you like me to keeping searching?”