Very good move. In my experience, for system programming at least, GPT 5.4 xhigh is vastly superior to Claude Opus 4.6 max effort. I ran many brutal tests, including reconstructing for QEMU the SCSI controller (not longer accessible) of a SVSY UNIX of the early 90s used in a 386. Side by side, always re-mirroring the source trees each time one did a breakthrough in the implementation. Well, GPT 5.4 single handed did it all, while Opus continued to take wrong paths. The same for my Redis bug tracking and development. But 200$ is too much for many people (right now, at least: the reality is that if frontier LLMs are not democratized, we will end paying like a house rent to a few providers), and also while GPT 5.4 is much stronger, it is slower and less sharp when the thing to do is simple, so many people went for Claude (also because of better marketing and ethical concerns, even if my POV is different on that side: both companies sell LLM models with similar capabilities and similar internal IP protection and so forth, to me they look very similar in practical terms). This will surely change things, and many people will end with a Claude 5x account + a Codex 5x account I bet.
GPT 5.4 is the surly physics PhD post-doc who slowly and angrily sits in a basement to write brilliant, undocumented, uncommented code that encapsulates a breakthrough algorithm.
Opus 4.6 is the L5 new hire SWE keen to prove their chops and quickly turn out totally reasonable code with putatively defensible reasons for doing it that way (that are sometimes tragically wrong) and then catch an after-work yoga class with you.
Who replies to you with fucking emoji brainrot
You are absolutely right!
GPT is also cautious and Defensive but opus is agreeable.
> and then catch an after-work yoga class with you.
That's cute, but do you mean something concrete with this, aka are there some non-coding prompting you use it for that you're referring to with that or is it simply a throwaway line about L5 SWEs (at a FAANG).
(FWIW, I find myself using ChatGPT for non-coding prompting for some reason, like random questions like if oil is fungible and not Claude, for some reason.)
It’s an analogy about the “personalities” of the models.
They are saying that Claude is more of a team player and conformist. It isn’t really much deeper than that.
I think the point they are trying to make is the golden retriever vibe/energy you get from Claude gives "after work yoga."
Thanks for confirming my impressions, it's been like 4 months now that I've arrived at the same conclusions. GPT models are just better at any kind of low-level work: reverse engineering including understanding what the decompiled code/assembly does, renaming that decompiled code (functions/types), any kind of C/C++, way more reliable security research (Opus will find way more, but most will turn out to be false positives). I've had GPT create non-trivial custom decompilers for me for binaries built with specific compilers (it's a much simpler task than what IDA Pro/Ghidra are doing but still complex), and modify existing Java decompilers.
Regarding speed, I don't use xhigh that often, and surprisingly for me GPT 5.4 high is faster than Claude 4.6 Opus high (unless you enable fast mode for Opus).
Of course I still use Opus for frontend, for some small scripts, and for criticizing GPT's code style, especially in Python (getattr).
Codex also gives you a lot more usage for $20/mon than Claude, so there’s not also that fear that high or xhigh reasoning will eat up all your quota. It really comes down to whether you want to try to save some time or not. (I default to xhigh because it’s still fast enough for me.)
In the SCSI controller work I mentioned, a very big part of the work was indeed reasoning about assembly code and how IRQs and completion of DMAs worked and so forth. Opus, even if TOOLS.md had the disassembler and it was asked to use it many times, didn't even bothered much. GPT 5.4 did instead a very great reverse engineering work, also it was a lot more sensible to my high level suggestions, like: work in that way to make more isolated progresses and so forth.
GPT 5.4 is remarkably good at figuring out machine code using just binutils. Amusingly, I watched it start downloading ghidra, observe that the download was taking a while, and then mostly succeed at its assignment with objdump :)
+1 to this, I've found GPT/Codex models consistently stronger in engineering tasks (such as debugging complex, cross-systems issues, concurrency problems, etc).
I use both OpenAI and Anthropic models, though for different purposes, what surprises me is how underrated GPT still feels (or, alternatively, how overhyped Anthropic models can be) given how capable it is in these scenarios. There also seems to be relatively little recognition of this in the broader community (like your recent YouTube video). My guess is that demand skews toward general codegen rather than the kind of deep debugging and systems work where these differences really show.
Or rather, it’s hard to ask everyone to side-by-side compare both products on their use cases. So the choice really comes down to word-of-mouth even though their use cases may be better served by Codex.
It's surprising to me how much LLM "personality" seems to matter to people, more than actual capability.
I do turn to Anthropic for ideation and non-tech things. But I find little reason to use it over codex for engineering tasks. Sometimes for planning, but even there, 5.4 is more critical of my questionable ideas, and will often come up with simpler ways to do things (especially when prompted), which I appreciate.
And I don't do hard-tech things! I've chosen a b2b field where I can provide competent products for a niche that is underserved and where long term relationships matter, simply because I'm not some brilliant engineer who can completely reinvent how something is done. I'm not writing kernels or complex ML stacks. So I don't really understand what everyone is building where they don't see the limits of Opus. Maybe small greenfield projects with few users.
> It's surprising to me how much LLM "personality" seems to matter to people, more than actual capability. > I do turn to Anthropic for ideation and non-tech things. But I find little reason to use it over codex for engineering tasks. Sometimes for planning, but even there, 5.4 is more critical of my questionable ideas, and will often come up with simpler ways to do things (especially when prompted), which I appreciate.
Aren't you saying here that the LLM personality matters to you, too? Being critical of you is a personality attribute, not a capabilities one.
Not necessarily. Criticism is the analysis, evaluation, or judgment of the qualities of something. This is a matter of intellectual act. However, you could say that being habitually critical can be partly a result of "personality" or temperament.
(Of course, strictly speaking, LLMs have neither temperament, "personality", nor intellect, but we understand these terms are used in an analogical or figurative fashion.)
> I'm not some brilliant engineer who can completely reinvent how something is done
With an honest evaluation of your own capabilities you are already far above average. Also its hard to see the insane amount of work that often was necessary to invent the brilliant stuff and most people can not shit that out consistently.
I use codex for cleaning after cloude and it always finds so many bugs, some of them quite obvious.
My non scientific tests has been that GPT models follow the prompts literally. Every time I give it an example, it uses the example in literal sense instead of using it to enhance its understanding of the ask. This is a good thing if I want it to follow instructions but bad if I want it to be creative. I have to tell it that the examples I gave are just examples and not to be used in output. I feel comfortable using it when I have everything mapped out.
Claude on the other hand can be creative. It understands that examples are for reference purposes only. But there are times it decides to off on a tangent on its own and decide not to follow instructions closely. I find it useful for bouncing off ideas or test something new,
The other thing I notice is Claude has slightly better UI design sensibilities even if you don’t give instructions. GPT on the other hand needs instructions otherwise every UI element will be so huge you need to double scroll to find buttons.
This is also what I noticed.
GPT doesn't know how to get creative, you need to tell it exactly what to do and what code you want it to write.
For Claude you can be more general and it will look up solutions for you outside of the scope you gave it.
I presonaly prefer Claude.
I think you might benefit from the "superpower" plugin. Add the word "brainstorm" before your prompt and it does a little bit better at figuring out how you want things.
What I like most about gpt coding models is how predictable of a lever that thinking effort is.
Xhigh will gather all the necessary context. low gathers the minimum necessary context.
That doesn’t work as well with me for Opus. Even at max effort it’ll overlook files necessary to understanding implementations. It’s really annoying when you point that out and you get hit with an”you’re absolutely right”.
Codex isn’t the greatest one shot horse in the race but, once you figure out how to harness it, it’s hard to go back to other models.
Yup I've mentioned this in another thread, I got gpt 5.4xhigh to improve the throughout of a very complex non typical CUDA kernel by 20x. This was through a combination of architecture changes and then do low level optimizations, it did the profiling all by itself. I was extremely impressed.
GPT5.4 with any effort level is scary when you combine it with tricks like symbolic recursion. I actually had to reduce the effort level to get the model to stop trying to one shot everything. I struggled to come up with BS test cases it couldn't dunk in some clever way. Turning down the reasoning effort made it explore the space better.
can you explain what you mean by symbolic recursion tricks in this context?
The model can call a copy of itself as a tool (i.e., we maintain actual stack frames in the hosting layer). Explicit tools are made available: Call(prompt) & Return(result).
The user's conversation happens at level 0. Any actual tool use is only permitted at stack depths > 0. When the model calls the Return tool at stack depth 0 we end that logical turn of conversation and the argument to the tool is presented to the user. The user can then continue the conversation if desired with all prior top level conversation available in-scope.
It's effectively the exact same experience as ChatGPT, but each time the user types a message an entire depth-first search process kicks off that can take several minutes to complete each time.
1000%. I have been running claude's work through codex for about a week now and it's insane the number of mistakes it catches. Not really sure why I've been doing this, just interesting to watch I guess.
Not to mention a billion times more usage than you get with claude, dollar for dollar.
It's widely reported that opus has been greatly reduced for a number of weeks since Mythos was released internally
Same for me, cf. https://news.ycombinator.com/item?id=47680123
The $100/mo giving access to GPT Pro (with reduced usage) is a nice counter to the just teased Claude Mythos. But GPT 5.4 xhigh being able to perform that kind of low-level reconstruction task is very impressive already.
I completely agree with you on both the technical and ethical reasoning.
Thank you for speaking out. I think it's important that reputable engineers like you do so. The Claude gang gaslighting is unhinged right now. It would be none of my concern but I have to deal with it in the real world - my customers are susceptible to these memes. I'm sure others have to deal with similar IRL consequences, too.