Hacker News

> 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.

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.

Calling this a "tupper ware" seems a bit emotional, you're intentionally disregarding many things that matter for devs in order to try to claim equivalence, rather than paying attention to the actual process of software creation.

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.

You're overestimating the extent to which individual developers have a choice here. My employer signed up for a Claude Code membership, I use Claude Code. I cannot use Codex.

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.

I can’t tell the difference between code written in vim or vs code but it matters substantially to the person writing the code. There’s stuff beyond just the output that goes into tool choice.

Ah that's always SO fun. It doesn't matter how "smart" the person actually are (or think they are) we are ALL susceptible to influence and blind tests are shockingly simple to implement.

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.

This is like saying you gave a Taylor Swift fan sheet music from 1984 and from Michael Jackson’s thriller and they couldn’t tell the difference.

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.

I think for developers the distinction is that ChatGPT is this commercial all in one solution for normies and Claude is specific for developers, in reality as you say the results for normal developers is indistinguishable.

The results are the same but I’ve found the process to get to the results are just more pleasant with Claude. I can’t put my finger on it. Overall most these models at the highest level are about the same in many respects but the UI/UX for some are just more enjoyable, for lack of a better term.

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.

I don't think it's marketing, for quite a long time Claude was clearly better and not everyone has adapted to the new reality where they have similar capabilities.

I don't think that's the only reason but you're spot on about OpenAI marketing being absolutely terrible. The primary product names of "Claude" vs "ChatGPT" highlights this remarkable difference. To the point where I'm seeing Claude completely take over the generic term for agent.

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?

1. It's 1 in 10 failures that can take half of your time or bugs that can take a long time to surface. Plus the way they change things largely depends on the current codebase (and how it was created)

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 picked Anthropic way early on, before Claude code even existed. Because they at least play lip service to behaving morally. That’s the most you can hope for these days really.

Everyone can be propagandised. It's a matter of pushing the right buttons.

It's a matter of what context is available to me at this time. I like LLMS. They improve my workflow to an insane degree. I think Sam Altman kind of sucks. I don't trust OpenAI. If they were the only kid on the block, I'd use Codex. It's entirely possible Anthropic sucks in the exact ways that OpenAI sucks but has better PR. I don't have time to deep dive to find out. I still like using LLMs. I started using Claude because Cursor, as a company, did something that I can't recall but gave me the ick. So I switched to Claude Code.

I still use Claude Code because I have the most experience with it now, and it's the harness that I understand on a granular level. If something comes along that is clearly better, or if it becomes clear the Codex is miles ahead, I'll try it and evaluate it. To your point, there doesn't seem to be much of a difference.

Arguing over this stuff feels kind of silly, like back in the day when my friends would give me shit for using mIRC instead of ircii or BitchX. I liked the GUI then because I did. I like Claude Code now because I do.

> I never want to hear from developers again that they are not susceptible to marketing.

Did you need to come to that conclusion?

Marketing has always been a significant part of new technology adoption. Whether it's for cloud adoption, for new programming languages, for new software development techniques, etc...

Steam and other game stores are pretty much the same but Steam is more popular because every one of their competitors has decided to continually shoot themselves in the foot over and over.

Even if Claude and ChatGPT were exactly the same, Claude would be more popular because OpenAI has decided to make some very unpopular moves and try to make money where popularity isn't required. At the moment that popularity still seems to matter.

I've always interchangeably used the models.

I don't look at benchmarks.

It's a non-deterministic tool. A lot of the shit going on with LLMs just doesn't make sense to me. All the tooling around like MCPs, they're all just putting stuff into context. So to me the tools aren't really robust and they make little difference.

Lots of AI psychosis going on these days. And I say that as somebody that hasn't written a line of code since Sept 2025

I don't think it's only marketing. OpenAI had the advantage of being first to the market, and in the beginning of the race it seemed that the future belongs to them. Then came the bad PR and unpredictable quality of their main product.

For general use, ChatGPT's answers have gotten worse over the last year. I abandoned it.

Been to an Anthropic event in Paris last summer.

They served caviar. It probably had good ROI.

The harness/UI that claude code brought was the thing that stole developer mindshare. Thats when people stopped coding in IDEs. Nothing to do with the underlying model.

Benchmarking 1 or a few samples isn't ever going to yield anything but noise. The actual benchmarks use thousands of tasks.

GPT 5.5 genuinely was back on top for a while there, but if you look at the past 2 years, being on Claude was better than being on OpenAI most of the time. If you're going to pick a tool and not switch constantly it was the right choice. Not to mention their tooling has always been ahead, and that gets ecosystem benefits.

Are they close and interchangeable today? Sure. But Sonnet was genuinely way better than anything OpenAI offered for a long time -- the valuation reflects that, not any given moment in time.

You confuse ease of using a tool with quality of output. A skilled carpenter can work both with high and with medium quality tools and prefer one over the other with no difference visible in the craft they produce.

> 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.

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.

I certainly can’t tell.

I honestly think I’d need weeks of all workday testing to even form an opinion… and some in depth training before that to use each given tool right…

And then … I might decide I can’t tell the difference.

As it is I use Claude and I don’t have the time to properly compare.

I find codex superior in speed and equal in quality, so it’s my preference. But Claude Code made prettier UIs last time I tested. Codex produces Microsoft-grade UIs. Very enterprise and ugly unless I actively steer it.

Pretty easy to tell depending what the code is. GPT follows this pattern is using maybe_something and using uppercase constants by default. Claude is a little more natural but tends to include more fallbacks than gpt5.5

Very similar thing happened when I was at a design event a couple of days ago. I’d say it’s even worse on the design end - there was a big discussion around how to optimize your usage of Claude. Not optimize your usage of AI, but Claude specifically, as it was the only model literally all of them were using. The biggest issue is they were all hitting their usage limits. I asked whether they had tried other, lighter models (Ie gemini or composer), and it was like I was speaking a foreign language.

Newer GPT (5+) models seem to forget imports more often than Claude and use all lowercase comments more (possibly as part of OpenAI's effort to make it more concise).

It also seems to use modern Java features like var and records more.

I found that the newest opus and 5.5 are definitely close enough where most of the work I do could be done with either. I've seen small differences in planning which I feel like Claude does do better, but I think both products are close enough where I wouldn't be upset if one disappeared.

The funny thing about Tupperware is that some of us have their products from many many years ago and they still work great.

I think we've had the same iced tea pitcher since I was 5 years old, for example. Solid.

Will we be able to say the same thing about Claude?

I use both, enough to reach Codex highest personal sub limits and Claude is stronger to me specifically because of how the flow of building feels. So the PR for any random task would be irrelevant to me.

It seems we're moving past the point where it's all about model capability. opus4.7 behaves better for me than gpt5.5 because I'm familiar with its idiosyncrasies. Sounds like you've got a good balance between them.

At the end of the day what matters is which team is better, not which model. If Anthropic continues to feel like the good guy, relatively speaking, then people are gonna chose to spend more time getting to know its products and less time with OpenAPI's and on average Anthropic's will be the more capable teams.

I think vibes are gonna matter more and more going forward. The potential for bad behavior on the part of an AI company is severe. We're gonna have to tolerate whoever we enable in this space, so I propose that we make their marketing teams work as hard as possible to show us which will supply better vibes.

> Couldn’t tell.

add deepseek v4 to it, and it will be close at 1/10 th the price. I use all three codex, claude, and deepseek, and they are close.

a) everyone is "susceptible" to marketing - so what

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

Agree wholeheartedly. I think that Anthropic has just invested more effort in creating a better DevEx than OpenAI, and so people just "feel" that claude code is better but they're about the same really, claude code might be 5% better at best.

> 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.

I think you're missing one (or more) of the facets individuals decide "better" is, for the subjective individual.

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.

edit: In short, it's the interactive UX just as much as it is the final output.

> I never want to hear from developers again that they are not susceptible to marketing.

It’s a really good signal of self-awareness/arrogance

Yes, which means that in the long run this looks ugly.

So much faith and money in this idea, and seeing how fragile it is, does not look good.

Instead of only hanging them evaluate the final output, you ought to also have a way to have them evaluate the process and agentic aspects in getting to said output. Claude Code outshines when you look at it end-to-end, in my experience.

So you black boxed a few 'success' test, while the main diference between the two is the way they get to the result?

So you both used Anthropic models (Opus 4.7 being from Anthropic)? I'm struggling to understand what your comparison really was here.

Modern Tupperware party. 100% agree! That’s the best framing I’ve heard in a long time!

Claude was the best for the longest time. GPT5.5 challenges that, but inertia is real

It's crazy hearing devs on this site claim Claude is 10x better than all other AI solutions. I think it is fomo. Claude $LATEST_VERSION is perceived as the best and anything else is "missing out". New version comes out? Suddenly the old version is worthless, how on earth did anyone get work done with that?

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.

The results may be the same but I personally find Claude nicer to work with. It seems to understand my intent better than GPT and needs less guidance. Maybe it’s just personal preference.

Honestly I have no idea how you couldn’t tell. Reading a PR I can see the difference without even reading the words. (I doubt I could spot the difference just looking at the code diffs though.)

Claude commit messages - well structured test plan, readable.

Codex commit messages - wall of text, no structure.

The big difference though is sitting with the tools and using them for work. These are for sure vibes, but I’m sure you could pull out metrics for # steering re-prompts for example.

Codex just goes off and solves the problem, usually comes back with a solve; Claude more often gives up or needs input. Opus gives a broader design discussion, better at conversation. Codex finds deeper/better edge cases.

I think it’s like EMacs vs Vim - you can get your work done with both. There may be some tasks where one is way stronger. A strict “Better” is quite hard to justify.

Ultimately tool choice is a mix of science and art/taste; I want to feel joy using my tools, and fun little pixel explosions make me happy. If a different tool makes you happy, that is also fine.

Should've used deepseek. That would have have been interesting.

A lot is changing. Like 9months ago, I was convinced Claude was best. I'm not so sure anymore :^)

for me personally it's two reasons:

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!".

Sure, none of this is rational.

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.

That comparison doesnt make much sense. The quality of a model isnt really its final output but the experience of working through a problem with it.

You sound upset at anthropic becoming bigger? Is this some kind of playstation/xbox fanboyism where people root for one company or another?

I like claude because it performs better in my experience, maybe ill check out codex to see if its good too, but i trust anthropic to have the sauce when it comes to coding.

The thing about developers claiming not to be susceptable to marketing is wierd. Im sure developers are susceptable to marketing and maybe there are some that claim to not be, but you seem to be overblowing that quite a bit, like anthropic ran a huge propaganda campaign to get claude out ahead. I think its not really conspiracy theory. Claude was just better at coding and the word spread among developers and people experience confirmed that. Maybe codex has caught up but people trust claude since its more established. Theres really no need to be mad about it unless youre invested in openai, financially or emotionally.

IME Claude has been a bit inferior. But, yeah, the marketing is just great.

If advertising is a multi-billion dollar industry then it has to be effective!

Which model produced code that ran faster, with less bugs, etc?

100%

The belief structures here are really interesting. Blind tests would likely illuminate a lot of why people think that

I mean, yes? Anthropic’s investors are seeing more upside now and valuing the company higher. Your thesis is that this additional value is driven by better marketing rather than a superior model. Could be! The truth is we’ll never really know with certainty what factors are doing the heavy lifting here, we can only guess and argue over who’s a better guesser.

Tribalism at it's worst. It's like the Coke and Pepsi comparisons from years past.

That sounds like someone desperately trying ton convince people Pesi is better than Coke

I think frontier models are bait now. I prefer a fast, less thoughtful model that adhere's to instructions. The code these latest models produce is often hot garbage and you still need to micro-manage. Fast with small chunks is better.

I find this pattern annoying and also commonplace: MY taste is correct. I AM right. The emergent properties of free market resulting from revealed preferences of free willed agents is WRONG.

Any name suitable to name this phenomenon?

Isn't the experience of interacting with the models appreciably different? It's not all about the outcome. Not to mention the harnesses are increasingly the real product.

in my experience out of the box Claude Code is the better tool if you want to spend 0 time on config

I have always found this field, especially in the last 10-15 years, to be incredibly fad driven to the point that it reminds me of things like fashion more than an engineering field.

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.

I don't think it's marketing, it's the "nobody got fired for buying IBM" effect applied to software developers choosing tools.

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.

I did a pair programming comparison over 3 month on Codex 5.2 and Claude Sonnet and my subjective experience was that based on cost and rollbacks to a previous commit Claude is significantly better. Especially in VS Code Copilot. I wrote a long Substack post about it. I would share its but its in the paywalled archive by now.

Maybe some of these companies will learn to stop appointing awful leadership then.

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.

This is especially true in a market with almost zero product differentiation.

Who doesn't think they are susceptible to marketing?

That seems like a strawman.

> Couldn’t tell.

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.

You're delusional. Anthropic's success is not only the model - it's the harness. I'd absolutely be able to tell the difference between Claude Code and, say, Gemini CLI.

Nothing new. I used to work with .NET and went to some meetups and conferences. There are some hardcore Microsoft fanboys out there. Didn’t even mix the kool-aid, ate it right from the packet. They only know MS products and seem scared of anything else.

Maybe not your typical HN crowd but marketing absolutely works on developers.

Sam Altman is so cartoonishly, over-the-top sociopathically shady that he makes JD Vance look like Benjamin Franklin. I mean, honestly, tricking third world people into retinal scans in order to get a scam crypto coin? Anyone using OpenAI for anything at this point should pause and examine their ethical compass.

If Claude meetups are the new Tupperware parties, OpenClaw meetups are something darker still. The tech is indeed worth discussion and celebration, but the Brands are clearly taking too much.

This is a really important insight. Great comment.

Claude has an "End Conversation" tool that it can trigger on it's own, forcing your interaction to a close based on it's own feelings towards the conversation.

I have no idea how this wasn't the end of Anthropic's positive public perception.

How can you say this as if supporting Dario is any better.

At the top level of anything there is almost no such thing as a non-asshole.

None of them care genuinely about you they just want your money.

Dario is genuinely as bad as Sam.

Altman does appear to be an asshole, but I have bad news for you if you think Anthropic are the good guys. If anything, they might be worse than OpenAI.

The idea that intelligence will be commoditized is completely counterintuitive. It comes from the belief that it can’t exceed our own. This is almost certainly not true at the limit. There will likely by many super intelligences like we see life in the wild

Sam and Dario are both assholes of the highest order. Sam just has more masks, and is better at wearing them.

>I think Sam Altman is an asshole and I prefer to spend my money elsewhere.

That's why he chose the OpenAI logo

I think all US companies are turning extremely anti consumer hostile thanks to unchecked unregulated capitalist greed and I would prefer my money to the Chinese underdogs. Cannot speak highly enough of Deepseek.

This is exactly it!

Sam Altman is the main perception problem for OpenAI. His background, history, trustworthiness, vibes/interviews etc are all negative PR when seen by the common man.

Dario is more knowledgeable, well informed, empathetic w.r.t. problems etc. In short, somebody who seems mature and trustworthy.

OpenAI and Anthropic are the same shit. That people think that Anthropic is in any way morally superior is sort of laughable.

He must have done something personally to you.

> Crazy. Those same enterprises will get sticker shock and leave.

They are already. Both for sticker shock, and also because of developer sentiment beginning to shift towards Codex. …and then in a month or two the winds will shift again, I'm sure.

It's interesting to see how Claude Code got commoditized so quickly.

---

> and they have the vastly better brand

Strong disagree there. Anthropic has pretty successfully branded themselves as the more ethical & 'human' of the two companies. (whether that's the actual reality is irrelevant)

I somewhat disagree, there was a major, major shift in developer sentiment towards agentic development starting with Opus 4.5 in October. Many teams started finding a lot more real value in Anthropic than they used to very recently. Things like OpenClaw are not a part of serious enterprises yet due to the security risk.

I’m not sure how anyone believes that per-seat pricing is halfway viable for AI, and I’m fairly sure the organizations I’m familiar with only REALLY started committing to spend after the shift to API pricing, due to the value they thought they were getting anyways.

OpenAI insider found!

My dude, this level of vitriol and hyperbole feels excessive. It's okay if you don't like Anthropic and you disagree with their business practices, but this is over-the-top even for the internet.

Anthropic hasn't done anything customer hostile to me.

I don’t know that most people care at all or even know about this. ChatGPT still far and away has the largest consumer market and brand recognition.

What Anthropic has done exceedingly well is work their way into corporations.

I have personally seen massive uptake over the last 6 months of regular people in corporations using Claude cowork. They are all genuinely amazed by what it can do for them.

OpenAI wants to be more of a Google. It’s increasingly seeming like consumer may not be as good of a play here

What's incredible is that Anthropic are clearly not saints but Altman makes them look like the good guys, reinforcing their marketing.

When Anthropic had the dispute with the Department of War over very meek conditions (a truly moral AI company would not be engaged in war crimes in the first place), it was a test for Altman, all he had to do was to take the same position. But because he's a psychopath he failed that very basic test.

I don't know, have you ever watched a video of Dario Amodei talking? I dislike them both immensely, but Dario both seems to take glee in scaring the shit out of people as a hobby with careless statements and his voice sounds like an impression of Elizabeth Holmes (the weird fake deep voice), and nothing he's said or done makes him seem more trustworthy.

Do you think Amodei is different?

Do you hold any amount of power in the world? A project that people care about, or a deliverable that someone depends on?

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.

Only on a good day?

Mechinterp in general is just completely undervalued right now (and agreed Anthropic's team is doing the most rigorous work, now accompanied by Goodfire). They're doing the closest work to neuroscience's in vivo 'thought-tracing', which is just the most wild science fiction sort of thing to be working on, and yet I feel the average person has no idea this sort of work is happening. When combined with the idea of the 'universal subspace hypothesis' (explored under the paper of the same name), you really start to bridge the gap from engineering to something more philosophical and spiritual. But I digress...

Did you also talk about "head and shoulders" and "pennant" patterns in stock charts? Or where the "smart money" is at? I'd like to subscribe to your paid newsletter.

Dario really gives: "I'll make the world a better place after I burn it to the ground, I promise."

I strongly believe the reason gpt-5.x performs so well on large projects is because of the focused training they've done on their dedicated apply_patch primitive.

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'm experiencing the same. Codex gtp-5.5 has more brilliant intuitions, write less code, i.e. it identifies the exact point in which the modification shall be done. Nevertheless, huge improvements on personality from opus 4.7 (it was too accomodating) to opus 4.8

GPT-5.5 is the better programmer but Opus 4.8 remains the better system architect and product designer.

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.

My problem with codex/gpt that is too verbose (mostly js and python): a lot of helper functions, a lot of 1 or 2 line functions used in 1 place only, a lot of types or proxy like objects.

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 ...

In what ways? LM Arena has Opus 4.7 w/ 1567 -/+ 7 vs. 1505 -/+ 10 from GPT-5.5 Codex in code. I'm currently using both.

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.

You're using last week's model; Opus 4.7 is old news. Opus 6.9 is the new hotness; it is a better product manager than GPT, and has more X productivity. It replaced our junior dev team, and tells me my hair looks good.

Opus 4.7 is not the current version of Opus.

Not everyone is a developer...

My experience as well. Although this week I've moved to Cursor and Composer 2.5. It's so fast that any faults can be iterated on super quickly. The model is just insanely good with code things.

source?

GPT 5.5 still invents facts rather than looking them up, and manages to come across both as condescending and sycophantic. It feels like talking to a used car salesman.

Yeah none of this is new if you've been working in this industry for a while

We've always been having these debates over whether my choice of tech is better than your choice of tech, same holy wars, different type of tech

The advice today is the same as it was 10/20/30+ years ago, pick what works for you and build something good with it

Nobody will actually care how you built it, regardless of whether its good or crap (although if it's crap you can blame your tools)

> jumping back and forth between companies

There is currently quite strong incentive to establish vendor lock-in for Anthropic and OpenAI. I see the ability to jump between companies quite important, especially for larger users. Right now it should not be hard but it can be much harder in the future.

OpenAI and Anthropic have the know-how for building much larger models that will be a lot smarter and run on datacenter-scale compute. This is a natural 'moat' that will be inherently hard to replicate for on-prem compute or small neoclouds running open-weight/local AI. They can easily coexist with a robust local AI scene.

> if in the future open weight models will be very near the quality of frontier labs, Anthropic and OpenAI will be out of business very soon

> Why would a business pay for Slack when IRC exists?

> Why would a business pay for Dropbox when FTP exists?

IMO bad take.

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.

this is like saying the car with the better engine wins, but all we're doing is commuting to work

I have the same impression. Strange to see this being downvoted & it was after reading the comment that I read the username to find out its antirez!

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.

This is an interesting claim to make. Up until quite recently (I mean that in the usual sense of the word, not the AI world sense of the word), almost no one had heard of Anthropic or Claude, despite an reasonably aggressive ad campaign, even at the point when most people would have known about ChatGPT.

Even now, I would guess that if you ask a normie off the street, they are far, far more likely to have heard of ChatGPT than Claude. Of course, Anthropic has been targeting businesses quite a bit harder than the general public for a while, so maybe that's not a fair test.

Anthropic inarguably does make an attempt at marketing their product. But I'm not convinced that the closing of the gap between them and OpenAI (as others have pointed out: I'm not sure it's defensible to claim that either is significantly ahead of the other given the paucity of available data, but they are certainly much closer than they were a year or two ago, when OpenAI was clearly in the lead) is mostly down to that. I think that, for a decent chunk of time (this one I mean in the AI world sense of the term), they had a very non-trivial lead in coding abilities. The developer and business world figured this out and jumped on board. That gap is largely now erased, but that's not enough to retake the momentum.

They just have no moral issues with spamming the internet with bots. They utilize blackhat tactics whenever they can to get an upper hand. Every social media platform is absolutely choc full of Anthropic and Claude promoting bots, and you know they're bots because they all repeat the same things, in the same wording. X in particular seems to have millions of them.

Except they raised $65 billion dollars, not $1.

Ai is overstated in my opinion but to hand wave the reality of them having created something that investors were happy to value at $1T is pretty unfair

They are selling shovels, not mining gold themselves though.

So it's more like selling a derivative on a promise to steal open source for you in a useful way.

> Actually, they're building the infrastructure for the next 50 years.

What infrastructure? The hardware would be outdated in 3 - 5 years, after all. What other infrastructure is needed for AI?

The dark fiber of our time?

By an order of magnitude.

tesla has been overvalued for almost a decade with little sign of slowing down, it really doesn't seem to matter anymore

"Investors who have poured hundreds of billions into closed-source labs are betting on an unprovable safety moat".

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.

they don't care, they're driving towards a cliff full speed and are all counting on jumping out at the right moment

My impression as well. OpenAI was riding the high of ChatGPT with a very confusing and seemingly unfocused offering beyond that. Anthropic was always laser focused on business use cases. Claude Code being the big one. Finance seems to be their next target.

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

My impression too.

Claude Desktop, Cowork, Code, Design all get meaningful new features week over week.

I can’t recall another vendor with such focus and velocity.

Google products are evolving at a glacial pace. OpenAI isn’t as focused on what knowledge workers need.

Gemini is probably one of the better AI's for programming, despite Google not making a decent official client to access it.

It is unclear which strategy will work in the end. 3.5 flash uses fewer tokens and is cheaper.

The "normie" market doesn't pay for enterprise features though. They might cost more in inference then they make back from advertising.

I don't see a problem as long as Chinese companies are not banned; expensive providers like Anthropic will prove that value is being provided, encouraging competition, which will take care of problems.