Modern tupperware party.
A colleague was convinced Claude is better so we played a game. We used the claude code and codex harness and I implemented some prs they needed with gpt5.5 and opus4.7 and asked them to identify which came from which only from the code.
Couldn’t tell.
Edit: i bet 99% of people here, if presented with a test where i gave 5 models but all of the results came from one, would not be able to discern this. Just vibes all the way down.
Anecdotally I hear of folks with workplace Claude Code subscriptions all the time. I'm not sure I've ever heard someone talk about their workplace Codex subscription. Anthropic clearly did a far better job chasing corporate customers while OpenAI was busy chasing consumers with Sora etc.
The test they (supposedly) ran with their coworkers to look at PRs from both is such a bad way to compare LLMs that I don’t think they’re very experienced with using them.
I'm sure they could also negotiate a similar deal with OpenAI but in my outsider experience it seems that negotiations around these kind of corporate contracts takes forever and when the selling point is "they're broadly pretty similar" I suspect the motivation isn't there.
> Couldn’t tell.
Why would you expect them to be able to recognize the signature of a model from a pair of PRs? I don’t understand why you think this is a useful test for anything when we have numerous benchmarks that run 100s of tests on models and both GPT-5.5 and Opus-4.8 perform similarly.
I have subscriptions to both. I run both on max reasoning. It is interesting to see the relative strengths and weaknesses of each model. You won’t always see it if you’re just scanning code. Some times one will spin for a long time on certain problems where the other has no problem finding the appropriate parts of the codebase and getting an efficient solution.
antirez made a comment that he and others found GPT-5.5 to be better at the optimization tasks he was working on than Opus. There are other classes of tasks where GPT-5.5 consistently stumbles where Opus will get a solution quicker. Lately I’ve been working on some code where neither model comes up with a good solution. That’s just how LLMs go.
The only reason you have seen more activity about Claude is that they got there first. Codex has been a step behind and GPT couldn’t match Opus at first. You’re testing them after they’ve closed the gap.
For example in your "test" you're only looking at output and ignoring the entire process of creation.
In addition to that process, you're ignoring that Claude Code was first and better for a long time, why would people switch for something that produces the same output? Claude Code has been way ahead in the process of agentic software creation for a long time, I still prefer its features. Even though I think that Opus 4.7 was a big step backwards, and I've been getting worse results seemingly every day with the churn of features at Claude Code, some of that may also be me testing the bounds of how little I can specify and still get acceptable results, so it's hard to know.
Calling all these concrete realities "marketing" is itself you trying to market Codex as "good enough" instead of paying attention to how we got where we are and where we will go in the future.
Software developers are the most susceptible of all population groups for amplifying their employers' new whims. There are true believers and useful idiots, but many are just mediocre and know that playing along will further their career for a couple of years.
In the end they will be fired anyway of course.
Codex I feel the need to be very specific and precise with. Claude… I feel like I can be lazy, which I enjoy.
Both still need to be reviewed stringently but I feel I can be more ambiguous with Claude and get better results than when Codex.
You definitely can in principle; that’s the entire point of the comment you are responding to. If one tool completes it in 10 minutes with little hand holding, and the other does it in one hour at 4× the cost and while needing a lot of steering, the former is arguably better even if the end result is the same.
Whether that’s specifically true and demonstrable of GPT and Claude is another question, but your blanket statement doesn’t hold as a general rule.
I think a more appropriate rephrasing would be 'You cannot simply make a claim that (model + harness) X is better than Y, but then have no discernible difference on dimensions you care about'. In the case of latest of claude code vs codex with gpt 5.5) both are similar enough in the dimensions people will care about in evaluating (vs. differing wildly in cost or time taken).
- which tool required more detailed goal-setting in the prompt?
- did one tool ask follow-up questions up front vs spread out over implementation?
- did either tool match existing coding styles?
- did either tool remind you about potential conflicts between what you asked it to build and other parts of the codebase?
There are a lot of ways to compare agents besides just the code. (Similarly, working engineers are not evaluated just on their code output.)
Sorry I think this misses the mark.
Because it's not the output but the process.
And sometimes the outcomes are not always discernable.
Codex and Claude are very different.
I use them for different things.
Their behaviour difference is obvious.
Of course it'd impossible for anyone to tell by looking at my code base 'how it was written'.
I've not used Codex to compare against, so I'm not claiming X is better than Y, but comparing tools simply on their output is naive.
That’s actually what my comment was based on; raw code output isn’t the only measure of quality. Engineers write better code if they have the tools they prefer.
Yup, like billions of capex. Unlike vim.
While there is no meaningful difference in the ability to write code, vim has earned it's reputation for having a learning curve. I'd argue that predisposition, that requirement for additional investment energy will bias the results towards attention to detail, and pure minimalism.
Convinced you can distinguish A from B? Ok! No problem, let's try! Can be at the dinner table for fancy wine or with agents, it's all the same, you try an option, another option, maybe all options from the same, and if you reliably can't tell well kudos, you are just like the rest of us!
It's easy to "know" in retrospect but blind test is where genuine difference can be found. Or not.
Over half of HN commentators visibly struggle to piece 3 or more complex ideas together.
How could anyone, who spent more than 30 minutes reading HN, expect otherwise?
I sometimes wonder how much of what I believe is bullshit I was fed through intentional propaganda. I do think as I’ve gotten older I’ve gradually identified and challenged some of it.
I have a strong affinity for Claude Code because of the interaction experience and overall tone / vibe / process. I am 100% willing to believe the code it produces is identical or possibly less good than Codex.
I enjoy working with Claude in a way I just don’t get from OpenAI. YMMV, you may feel just the opposite. But it’s a mistake to look at the produced code as the only dimension of these products.
There should be a material difference between the tools.
There is.
vim / emacs / jetbrains - different tools to produce code.
Codex and Claude are different.
It is like the employee who is slightly worse but is a brownnoser getting promoted more often.
And what do you know, that is what is happening. It is like the coke commercial with the nice music and beautiful person in the back.
Speaking of which, remember Pepsi Challenge? Coke lovers are like the claude code lovers.
And it really depends on the task. Is it a typical well defined bug, or is it simpel CRUD. Or does it require research, combining different sources of data in a complex and creative ways.
This is also why benches never show reality, and the only real understanding comes if you actually try to build something.
Repeat after me:
_Other people can experience things you do not experience and it is still valid, and not a delusion_. They are not sheeple who fell for marketing.
FWIW most of the normies I know are using Claude
This has to be in some far side gallery somewhere
I do think OpenAI is doomed due to bad leadership. What you said (that the marketing is relatively terrible) and what others are saying here (that the product is worse) is damning isn't it? Are they really failing on all fronts?
Same reason people buy the RTX 4090 and 5090 cards - overpriced but they must have the "best". Never mind the diminishing returns trying to max out PC settings (3-4x performance hit for an almost imperceptible increase in graphics, ignoring DLSS) - it's the psychological cost of having to move a slider down a notch.
I've been using Google and now DeepSeek v4 and I am having absolutely no problems and it's a fraction of the cost. I'd love for Claude to be 10x better but it just isn't, for my use case anyway.
Vibes and tribalism will prevail until one of emerges as clearly and unambiguously superior to the other.
I think it’s great, but coming from Claude Code it did feel like going back in time by ~6 months in model capabilities. This isn’t a big deal to me for what I do, but the difference is definitely there.
I’m being pedantic/splitting hairs, though. I’ve obviously switched to DeepSeek full-time because it makes more sense to me pragmatically — I spend a few more tokens to get the outcome I want, but the tokens are cheap as dirt and the API is faster.
Perhaps I should plug it into Claude Code and see how it performs? I haven’t tried that.
2. In my case codex seem to be writing a more solid code, but I still use claude most of the time because it's my witty rubber ducky and I can actually sometimes force some legit insights out of it. Codex is much worse at this. And whether that matters or not depends on the project.
I think you're missing one (or more) of the facets individuals decide "better" is.
Early on i hopped between all the providers. Code quality for SOTA at the time was pretty decent if you didn't ask it to solve challenging problems. However the thing i found most difficult is consistency in how it listened. Eg Gemini (i forget what version, not current) was super prone to focusing solely on the functionality/goal, but not any of the directions on how to write the code. It would throw in comments everywhere, document in a manner i didn't want, use abstractions i told it not to, etc.
How well a model would follow instructions to drop their horrible "isms" was the #1 criteria for me. If i have to constantly remind the model not to do X behavior then it's a terrible model.
With that said, that is why i chose Claude for the last N months. However i've stuck with Claude because dealing with these "isms" and their little behavioral nuances is a chore in itself. I've found you have to learn the model just as much as anything, and so the idea of hopping these days when i'm just trying to get shit done is not likely.
These days for me personally, Claude has to give me a reason to switch rather than me investing even more money (i'm on the 20x plan) in other providers. I'm definitely not committed to Claude Code, but i am tired of the LLM churn, tooling churn, subscription churn, and the general fear of which providers we can trust.
This is complicated by the way that the coding agents inject prompts that preempt and potentially undermine user instructions. I suspect that one of the reasons Codex works way better for me than Claude Code in certain projects is that the latter adds some garbage like "go ahead and write repetitive copy/paste code, keep it simple, take shortcuts" to every session. A fair test would have to hide but more or less still use the harnesses, not just the models.
The Agent SDK uses the Claude Code harness. I went through this exercise recently to see if we could cut down the system prompt to save some token usage.
Lots of fat to trim, but nothing about writing repetitive copy/paste code. There are like 3k tokens describing the barely functional memory system that you get to pay for every single turn.
For general use, ChatGPT's answers have gotten worse over the last year. I abandoned it.
I don’t think that applies to most on here tho.
Edit: Oh they’re trolling, nm. :-/
b) therefore a preference for Claude is marketing - complete bollocks
Either the tasks you chose were well below the capabilities of top models, or meaningful differences for preference are elsewhere, or both.
Your comment is probably energy-efficient and sustainable, however, because you could use it again and again when another comparison comes up, like Vim vs Emacs, or tea vs coffee
Some of its timing: Claude Code was good before other harnesses and so behaviors (and contracts) were timed to lock in on that ecosystem.
Some of it was ethical/political: Anthropic fighting with the Trump admin about use of the model.
Some of it is social: Never overrate a CEO just being kind of perceived as a piece of shit by people who have power to influence decisions.
But switching costs are low! Because of the same models!
Let the race to the bottom commence. Hopefully before the monopoly/collusion starts.
So much faith and money in this idea, and seeing how fragile it is, does not look good.
They served caviar. It probably had good ROI.
That seems like a strawman.
Google came pretty close at times
1) Brockman ($25M) and Altman ($1M) both personally donated to Trump/MAGA.
2) Anthropic pushed back against DOD's demand for unrestricted use of AI to kill people while OpenAI eagerly said "please use ours!".
I think OAI actually legitimately increased p(doom) for us all. Very strange behavior for a company that is supposedly concerned about x-risk.
It’s one of the things I don’t like about it. All humans are susceptible to herd behavior and influence but engineers should be at least a bit more hard nosed and reason more from first principles.
It's the same reason why most of the software out there keeps using bloated technologies that are most of the time the wrong fit for the product.
And the same applies to tooling. Nothing new.
Having a sleazy CEO like Sam Altman or Elon Musk is a business risk. Many potential customers don’t like these people and they say abrasive and alienating things publicly.
Rolling over to the DoD’s desire for fully automated weaponry is more bad marketing. How many people switched from OpenAI to Anthropic over that? I sure did. Anthropic’s willingness to burn that bridge over an ethical stance said a lot about the company to me.
I’m not going to use OpenAI products for these reasons among others.
I’m also not going to use Cursor as xAI plans to acquire Cursor.
Maybe it’s foolish of me to avoid those companies for such petty reasons, but that’s not my problem. That’s their problem.
It takes years to build trust and hours to burn that trust to the ground. Customers can hold grudges for a lifetime.
I can tell. It's night and day.
Last year I used a bunch of models to try to generate Rust code. They all sucked.
This February I tried again and used Claude to generate Rust code. I have never been more stunned in my life. It's just as good as I am, and 30x faster. No fluff, the code is verbatim just as I would have written.
I then tried other models. Total disappointment.
I've continued to repeat this experiment. Opus is the only model that can write Rust reasonably.
Codex produces junk to this day. It passes variables that aren't needed, it abuses pointers, it creates overly verbose monstrosities...
I don't want any single company to win. I want OpenAI to be competitive. I want open source models to win. But right now, Claude Code and Opus are it.
Having looked at a bunch of known or suspected (based on the intent of the code and/or what I know about the developer(s)) LLM generated rust, there's only a few explanations here:
1. You're way better at prompting than (virtually) anyone else.
2. You're vastly overestimating how good the rust code it produced is.
3. You handheld the model throughout and made lots of edits.
4. Your hand written rust code is very bad.
Because from every example I've seen, these models write horrible rust. Sure, it may technically pass all the tests, but it's horribly pessimized, badly organized, doesn't even attempt to use the type system, if there aren't bugs now there will be the second it tries to refactor or add a new feature, etc. etc.
(I also strongly suspect that the same would be true for other languages, but I can detect it in rust more easily because it's my main language)
I have no idea how this wasn't the end of Anthropic's positive public perception.
Frontier models being commoditize is inevitable. OpenAI thinks they're still competing on technology, and not user experience and market reputation otherwise they'd understand the continuous negative PR generated by Altman's chaos is going to cost them everything.
Just curious how you can afford to care about the guy 7 levels above the men that built and support the API that you buy.
People can spend money how they wish. SamA is a prick, so I don’t buy from his company. I don’t buy from Microsoft or Oracle either. Giving a company your money is explicitly supporting them and everything they do. Are you going to force me to buy products from people I don’t agree with?
Some don't, and find it hard to believe others really do.
Not saying that’s right or wrong, but it’s clearly a factor holding OpenAI back at this point.
They'll kill us all, or they'll kill each other. They sure as hell ain't making the world a better place, like they promised.
Upshot - poetry expertise does not seem to be the primary focus these days, perhaps to the detriment of the entire world. We did move on from training scaling to “test time” scaling (which I hate as a name btw), Ilya does not seem to have been needed, (although I am really curious what he’s building).
My prediction that you want to be deeply embedded and really rich and part of global infrastructure feels good. My suggestion that oAI / MS would be able to use the lead in 2024 to extend was wrong.
Neither of us talked much about coding as a product that would drive value and behavior, which is super interesting to me, we were probably six months from seeing real competence of any sort there way back in June 2024.
We both seemed to think there would be a single breakout company, or could be one, (although I did suggest buying the basket), clearly not the case with GOOG oAI and Anthropic all posting serious revenues this last quarter / year.
One area of Anthropic that was nascent in 2024, but that I have come to think is super valuable is their mechinterp group. I still don’t see work done by other labs (at least published) to nearly the quality of Anthropic. And the group has clearly moved into a period of productivity; there’s a good chance in my mind it could provide a truly enduring strategic advantage as a tool to be used by the taste makers steering the ship. In 2024, interpretability seemed almost impossible to get a handle on — today, the sustained chipping away at the problem makes a lot more look possible.
The official implementation of apply_patch is well thought out. It is a two-phase process that will not actually make any changes until all files in the change set are not ambiguous. The pre-commit error feedback usually fixes anchoring issues with one or two additional attempts. It generally goes something like:
Reading file A L1:154
Reading file B L1:123
Attempting to apply patch...
[anchor errors for both A & B]
Reading file A L43:67
Reading file B L50:74
Attempting to apply patch...
Patch succeeded! Running compilation & unit tests...
Techniques that replace the whole file or depend on find-replace are useful in more isolated contexts. However, when you need to refactor 20+ files, something like apply_patch is what you want. Anything that depends on specific line numbers for actual replacement targets is a total dead end for complex edit scenarios.
https://developers.openai.com/api/docs/guides/tools-apply-pa...
I have specific skills for trying to avoid this, but nevertheless I spent half of the time fighting with its verbosity.
Currently, I'm trying to scaffold the functions/classes I know I need with NotImpelmented and ask it to implement only inside those specific places. It's a little bit better, but I still have to fight with function in functions definitions ...
Codex is very "miss the forest for the trees", but is much better at successfully making large changes in large codebases. Claude Code makes more mistakes, but has more taste and a better grasp on idiomatic and elegant software development.
If you can afford to, I recommend juggling both.
This is not a jab, but a genuine curiosity of mine.
But I feel like an expert who can drive GPT aggressively will out perform Opus. It’s why some smart people I know are opting for GPT and have fallen off on Opus. It’s like asking an F1 driver to sit in a taxi.
Admittedly my recent experience tilts Opus now 4.8, but you and others have my interest piqued re: GPT-5.5 Codex so I'm trying that more now.
As far as its tone... Both feel like sycophantic as hell to me. To be honest, they just all feel so.
So does Claude, what’s your point?
I used it and ChatGPT this week in trying to assist troubleshooting a complex DB related issue and Claude had to apologise no less than three times in which it admitted to talking complete shit.
Just one example of the kind of shit it dribbled:
> I need to be upfront with you. I should not have claimed X as if I knew that for a fact. That was overreach on my part.
Of course every AI company has been over promising and pumping the numbers as much as possible but OpenAI has been hitting the reality wall more because both their people not being able to keep improving at a faster rate and their whole cost structure and financial plates spinning.
This doesn't invalidate the fact Anthropic is also overhyped to the max for their IPO.
Stealing peoples tokens because you use a product they don't like... That shows the morals they have. Actions speak louder than words. Disabling peoples caches because they disable telemetry was another juicy one that I don't believe is on this site. In fact there are far more I remember that aren't even listed here.
The chutzpah is remarkable.
So it's more like selling a derivative on a promise to steal open source for you in a useful way.
You can theoretically do most things AWS does most of the time, yet people pay premium for it and keep paying for it, even though alternatives are cheaper, simpler and more performant.
I'd bet you that after 20 years OpenAI and Anthropic would still be around and kicking.
You might have a subpar product (for the price) but the reputation and history is what makes people open their wallets.
Now, I think that with these companies IPO'ing and Nasdaq and other bending themseleves and their rules to cater to them (as in case of SpaceX), these companies are very close to an IPO.
So for the employees, they are probably gonna get good evaluations, atleast in the short term and perhaps they are having a problem which is worth having.
But as you have suggested, I feel like the whole thing might be flaky especially given open source models. I believe that OSS models are at worst close to literal SOTA ~6 months ago.
So OpenAI & Anthropic have to somehow always be on the edge to get better models to not lose this (imo) very small time grip that they have, all while losing billions of dollars and having to worry about profitability & so many other concerns in it of itself.
I don't think that there is any other thing inside CS or any industry where two pieces of software being almost comparable enough with not much moat around except a diff of 6 months best, is something on which trillions of dollars float around on. We don't know how things will pan out but if I have to guess, It might not be looking good for OAI, Anthropic over especially the longer horizon.
Nobody is investing in closed-source labs for safety reasons, being able to explore more in details what and how the model is thinking is nice but by no means a game changer. What matters to investors and most of the users is that the model gives the right answer at the end.
> The new valuation is nearly three times higher than the company’s February valuation, when Anthropic was estimated to be worth around $380 billion.
> In March, OpenAI was valued at $852 billion following a record $122 billion funding round.
Basically, today (Late May) we're declaring Anthropic the most valuable. They've nearly tripled in value since February. But also, OpenAI was $852B in March and presumably has grown since then.
In a few weeks we'll either have a new rounding of funding for OpenAI or they'll announce their IPO and the hype train will be abuzz that they're now the most valuable.
OpenAI. Spent its resources on AGI whilst Claude worked on making programming work.
Google Gemini is out of the race entirely its programming AI is a joke.
Like actually iterating hard to make them useful. Many, many details matter here.
I haven't tested the similar OpenAI/Google tools in detail lately though. Previously I found them way too generic and unpolished to be useful.
Is there something to this?
Anthropic has much narrower capabilities. No image generation, no video generation, no 3d world models, barely any voice stuff. But they know who their target customers are, and their API has a model selection anyone can understand and pricing that rarely changes. Focus and predictably