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.
> 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.)
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).
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.
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.)
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.
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.
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.
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.
It's interesting seeing all the ChatGPT users in this thread, knowing what we know about OpenAI. Either they don't care about what OpenAI does, don't know their reputation, or feel like their use is too insignificant to matter.
This is sarcastically-stated but an excellent point, and an honest answer will come up with a vanishingly small list. We geeks may think we care about Important Things, but our industry cares for nothing but money and power — morality is a hindrance to the accumulation of those.
If an AI company has done unethical things do you think it is inappropriate to discuss that? Take Grok: among other things it created sexualized images of underaged women without their consent, not by accident but as a feature. Is that just something you want to ignore? In response the people in charge merely restricted the feature to paid subscribers instead of removing it.
Do you think people who mention grok creating CSAM is a holier-than-thou attitude? Do you not think the people who ignore that are worse than other people?
I assume in any sort of thread on a topic like this there is going to be inorganic activity. These companies are all fighting rather hard to try to gain marketshare, potentially worth $trillions, with a product fully capable of producing endless reasonably compelling content to populate an account, a website, or any other basic proof of identity one might ever want.
It's probably never been the case that plurality of views meant anything since online is a bubble to begin with, filtered by endless biases wherever we happen to be reading, making it an even more fringe bubble, but the advent of AI has pushed it all over the edge to the point that perceived pluralities are just completely and utterly meaningless. Somewhat depressing for a one who enjoys online chat as a pasttime, but it's the reality of the world now.
What's a good way to think about this? Because it does cross my mind about the billions of dollars at play - at the same time, I'm not a pessimist. I think my middle ground is kind of just the usual, taking things with a grain of salt. I mean, I chose to reply to this comment in good faith it's human to human.
I hope I keep that faith. I hope our billions of neighbors on the web enable me to keep that faith over the coming years. Definitely uncertain about the future of the web but want to love it like I've loved it 1990s-today. (Guess I should volunteer w/the EFF while job hunting, try for for-purpose jobs...)
5.4, in my own testing, was almost always ahead of Opus 4.6 for reviews and planning. I'm on plus plan on openai, so I couldn't test it so deeply. Anyone who had more experience on both could perhaps chime in? Pros/cons compared to Opus? I'm invested in Claude ecosystem but the recent quality and session limits decrease have me on the edge.
Same for me. I'm on $20 plan for both and I use them both interchangeably. Similar "intelligence" imo. Just different way of doing things, that's all. But Claude is getting worse in terms of token usage so I've cancelled my plan last month.
Yeah it's probably a bit better overall. 5.4 is a month newer than Opus 4.6
My guess is that 5.5 will come out soon and be significantly better so you'd want to be using Codex then, but then when Opus 5 comes out probably back to claude code
Also 5.4 has fast mode, and higher usage limits since it's cheaper
Just curious as I've often heard that Claude was superior for planning/architecture work while ChatGPT was superior for actual implementation and finding bugs.
Claude makes more detailed plans that seem better if you just skim them, but when analyzed, has a lot of errors, usually.
It compensates for most during implementation if you make it use TDD by using superpower et al, or just telling it to do so.
GPT 5.4 makes more simple plans (compared to superpowers - a plugin from the official claude plugin marketplace - not the plan mode), but can better fill the details while implementing.
Plan mode in Claude Code got much better in the last months, but the lacking details cannot be compensated by the model during the implementation.
So my workflow has been:
Make claude plan with superpowers:brainstorm, review the spec, make updates, give the spec to gpt, usually to witness grave errors found by gpt, spec gets updates, another manual review, (many iterations later), final spec is written, write the plan, gpt finds mind boggling errors, (many iterations later), claude agent swarm implements, gpt finds even more errors, I find errors, fix fix fix, manual code review and red tests from me, tests get fixed (many iterations later) finally something usable with stylistic issues at most (human opinion)!
This happens with the most complex features that'd be a nightmare to implement even for the most experienced programmers of course. For basic things, most SOTA modals can one-shot anyway.
The title is misleading. The only thing they seem to have done was add a $100 plan identical to Claude's, which gives 5x usage of ChatGPT Plus. There is still a $200 plan that gives 20x usage.
It's not worse, Anthropic simply has no equivalent model (if you don't consider Mythos) of GPT 5.4 Pro. Google does though: Gemini 3.1 Deep Think.
GPT 5.4 Pro is extremely slow but thorough, so it's not meant for the usual agentic work, rather for research or solving hard bugs/math problems when you provide it all the context.
So, reading the tea leaves, they're either losing subscribers for the $200 plan, or they're not following the same hockey stick path of growth they thought they were... maybe?
Edit: I wonder if this is actually compute-bound as the impetus
Nope, it's just that a lot of people (especially those using Codex) asked us for a medium-sized $100 plan. $20 felt too restrictive and $200 felt like a big jump.
Pricing strategy is always a bit of an art, without a perfect optimum for everyone:
- pay-per-token makes every query feel stressful
- a single plan overcharges light users and annoyingly blocks heavy users
- a zillion plans are confusing / annoying to navigate and change
This change mostly just adds a medium-sized plan for people doing medium-sized amounts of work. People were asking for this, and we're happy to deliver.
If you want it to deeply research something pro is great. I had a problem I just couldn’t find with my oven so I gave it a lot of information and it went off on its own for about 2 hours and then gave me what I needed to fix the problem (fan was turning off too quickly which was causing the panel to overheat). I have no idea how it figured it out and I couldn’t find anything after hours of googling so it was very impressive. I even went and googled for it once I knew what the problem was and I still couldn’t find the solution that it came up with.
Thanks for sharing this experience. Does it cost a lot of token in the deep analysis - which will make the $100 plan much quicker to drain all budgets.
I think it’s gonna be very hard to blow through your tokens just using chat. I mostly bought the plan so I could use Codex and on the $200 a month plan I’ve basically been using it 15 hours a day almost nonstop and I don’t run out of tokens for the week.
Notably, up until now Pro had 6x usage of Plus. So the title is only slightly misleading.
On the other hand, the benchmark of Plus usage seems to be to be all over the place, so it’s difficult to say now how does the usage compare to the old Pro.
That has me quite tempted. In general, I stay under the Plus limits, but I do watch my consumption. I could use `/fast` mode all of the time, with extra high reasoning, and use gpt-5.4-pro for especially complex tasks. It wasn't worth 10x the price to me before, but 5x is approachable.
For my money, on the code side at least, GitHub Copilot on VSCode is still the most cost effective option, 10 bucks for 300 requests gets me all I need, especially when I use OpenAI models which are counted as 1x vs Opus which is 3x. I've stopped using all other tools like Claude Code etc.
I use both GH Copilot as well as CC extensively and it does seem more economical, though I wonder how long this will last as I imagine Github has also been subsidizing LLM usage extensively.
FWIW it feels like GH Copilot is a cheaper version of OpenRouter but with trade-offs like being locked into VSCode and the Microsoft ecosystem overall. I already use VSCode though and otherwise I don't see much downside to using GH Copilot outside of that.
You’re not locked into vscode. There are plugins for other IDEs, and a ‘copilot’ cli tool very similar to Claude Code’s cli tool.
I also wouldn’t say you’re locked into Microsoft’s ecosystem. At work we just have skills that allow for interaction with Bitbucket and other internal tooling. You’re not forced to use GitHub at all.
I'm hopeful because Microsoft already has a partnership and owns much of OpenAI so can get their models at cost to host on Azure with they already do, so they can pass on the savings to the user. This is why Opus is 3x as expensive in Copilot, because Microsoft needs to buy API usage from Anthropic directly.
Sonnet is the worse model though, therefore it's expected that it is cheaper, the comparison would be Opus and GPT. That Anthropic's worse model is the same request cost as the best OpenAI model is what I mean when talking about Microsoft flexing their partnership.
I tried Claude Code for a week straight recently to see what all the hype was about and while it pumped out a bunch of reasonable looking code and features I ended up feeling completely disconnected from my codebase and uncomfortable.
Cancelled the plan I had with them and happily went back to just coding like normal in VSCode with occasional dips into Copilot when a need arose or for rubber ducking and planning. Feels much better as I'm in full control and not trusting the magic black box to get it right or getting fatigue from reading thousands of lines of generated code.
Anyone who says they're able to review thousands of lines effectively that Claude might slop out in a day are lying to themselves.
I don't like calling a posture 'ignorant' , but I think that's what we have here. I don't mean that as an insult.
It's likely you didn't learn how to use the tool properly, and I'd suggest 'trying again' because not using AI soon will be tantamount to digging holes with shovels instead of using construction equipment.
Yes - AI can generate slop, and probably too many Engineers do that.
Yes - you can 'feel a loss of control' but that's where you have to find your comfort zone.
It's generally a bad idea to produce 'huge amounts of code' - unless it's perfectly consistent with a design, and he architecture is derived from well-known conventions.
Start by using it as an 'assistant' aka research, fill in all the extra bits, and get your testing going.
You'll probably want to guide the architecture, and at least keep an eye on the test code.
Then it's a matter of how much further 'up' you can go,
There are few situations in which we should be 'accepting' large amounts of code, but some of it can be reviewed quickly.
The AI, already now in 2026 can write better code than you at the algorithmic level - it will be tight, clean, 'by the book' and far lesss likley to have erros.
It fails at the architectural and modular level still, that will probably change.
The AI 'makes a clean cut' in the wood, tighter to the line than any carpenter could - like a power tool.
A carpenter that does not use power tools is an 'artisnal craft person' , not really building functional things.
This is the era of motor cars, there is really no option - I don't say that because I'm pro or anti anything, AI is often way over-hyped - that's something else entirely.
It's like the web / cloud etc. it's just 'imminent'.
>Anyone who says they're able to review thousands of lines effectively that Claude might slop out in a day are lying to themselves.
The amount you can review before burning out is now the reasonable limit, for the same reason that a car is supposed to stay at the speed you can handle and not the max speed of the engine.
Of course, many people are secretly skipping reviews and some dare to publicly advocate for getting rid of them entirely.
> For the same reason that a car is supposed to stay at the speed you can handle and not the max speed of the engine.
As we know with driving, sensible drivers stick to the speed limit most of the time, but there's a good percentage of knuckle draggers who just love speeding, some people get drunk, some they just drive the wrong way down the highway entirely. Either way it's usually the sensible people who end up suffering.
That's my point - it's great as a tool to talk something through or rubber duck it, but as soon as you just let it loose to slop out thousands of lines a day and never read them all you're really doing is filling your base with thousands of lines of technical debt.
I think most people doing what is now called agentic development, aren't following most established dev methodologies and are to a great extent playing it by vibe.
The codebase disconnect is real.
We are like blue collar workers that need to hit the gym to maintain the body that our cavemen ancestors could maintain by doing their daily duties.
Of course it is. Returns are diminishing, AGI isn't happening with current techniques but it is good enough to sell, so it's time to monetize. I just got an email from OpenAI as well about ads in their free tier (I signed up once out of curiosity).
Yes, and now they (will) have ads in the lower tiers and eventually will add them in higher tiers as well. It's no different to Netflix's methods. Therefore the statement would have been proven right if it were claimed back then.
You can't just say because they've added more things the old things are over - the old things actually have to go away first. Eventually they may get there (or not). It may be another few years (or not). Nothing is actually now over though any more than it was now over in 2024.
Additional tiers to get people to move up as they enshittify the lower tiers. We already see it in other companies as well as OpenAI themselves so my inference is based on that, not to wait and see until they do indeed enshittify it.
Both Pro plans include the same core capabilities. The main difference is usage allowance: Pro $100 unlocks 5x higher usage than Plus (and 10x Codex usage vs. Plus for a limited time), while Pro $200 unlocks 20x usage than Plus.
So curious that the cost in the comparison is just a flat $100, not "$100 or $200" and yet the usage has the "or". Surely just a lapse in copy editing.
Anthropic is the exact same way, I think they're just trying to avoid having 5 different subscription tiers visible. Probably needing 20x is very niche
Suppose you enter a casino and the owner welcomes you in and sees that you are a frequent loyal s̶p̶e̶n̶d̶e̶r̶ customer (with the amount of tokens you are spending a month) with an existing membership.
With this new VIP membership that comes with 5x or 20x usage, if you spend $100 you get 5x. $200 you get 20x and you get to spin the wheel and use the slot machines unlimited times even at peak hours more than most without any restrictions, 24/7, no waiting for hours with priority.
So spend more to get more abundance and more simultaneous spins at the wheel.
Except if you're trying to abuse the slot machines themselves or sharing or reselling your membership to other customers who want a spin at the roulette wheel; but were previously banned. [0]
They are actively exploiting the compute shortages of Anthropic. In our team we're pushing for more or less vanilla and portability, since the best harness today might not be the best one in 6 months.
For me it's not the price. It's the fact that they obviously read my prompts and may even use a derived version of my data for training. As it's very clear in the meantime that SAMA lies most of the time, there's just no way I can trust this company in any way.
It really feels like LLMs will mostly become tools for tech workers rather than the kind of civilization-level transformation sama has been peddling. Every single comment here seems to confirm the above.
The missing piece isn't smarter models, it's making them usable for normal people. OpenClaw proved there's massive demand for AI agents but the setup and maintenance is brutal. We built Atmita (https://atmita.com) from scratch, not based on OpenClaw, cloud-native so there's nothing to install. It can hit any REST API directly, integrates with 800+ services out of the box, has a built-in browser, runs scheduled tasks in the background, and has built-in approval controls. All through a chat interface anyone can use.
This is kind of the goldilocks zone for LLMs right now.
I wouldn't mistake this for any kind of capability plateau. There is a massive push towards making transformers the engine of humanoid (and other kinds of) robotics, we just haven't reached the hype moment for those yet.
> I wouldn't mistake this for any kind of capability plateau. There is a massive push towards making transformers the engine of humanoid (and other kinds of) robotics, we just haven't reached the hype moment for those yet.
Problem is that the fuel to get this train going relies on investors money. Investors aren't going to be happy with the quote I took from your message.
And that's the real bet really, can the industry turn the spark into fire before the investor money runs out?
> the quality differential versus the competitor is night and day.
This myth about the inferiority of ChatGPT and Codex is becoming a meme.
I have active subscriptions to both. I am throwing at Codex all kinds of data engineering, web development and machine learning problems, have been working on non-tech tasks in the "Karpathy Obsidian Wiki" [1] style before he posted about it.
Not only does Codex crush Claude on cost, it's also significantly better at adherence and overall quality. Claude is there on my Mac, gathering dust, to the point I am thinking of not renewing the sub.
There are plenty of fellow HNers here who feel the same from what I read in the flamewars. I suspect none of us really has a horse in this race and many are half-competent (in other threads, they mention they do things like embedded programming, distributed DL systems, etc.)
I'm starting to suspect a vast majority of people pushing the narrative that Claude is vastly better haven't even tried the 5.3 / 5.4 models and are doing it out of sheer tribalism.
I have access to effectively infinite API tokens for all models from Anthropic as well as OpenAI. The differential in performance in complex tasks is vast and strongly in favor of Opus, in my experience. I do not use the official harnesses for either model, though - as they are not my taste.
Codex is closer to my taste, as it is at least a native app and not typescript slop. But the model is just not up to snuff.
Disagree. I use codex extensively. It just works so well with vscode and python. Claude with ridiculous limits - thanks no. For some even xAI is good fit.
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
> 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.)
GPT is also cautious and Defensive but opus is agreeable.
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).
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.
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.)
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.
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.
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?
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.
It's interesting seeing all the ChatGPT users in this thread, knowing what we know about OpenAI. Either they don't care about what OpenAI does, don't know their reputation, or feel like their use is too insignificant to matter.
What has the tech industry ever resisted on moral or reputational grounds?
This is sarcastically-stated but an excellent point, and an honest answer will come up with a vanishingly small list. We geeks may think we care about Important Things, but our industry cares for nothing but money and power — morality is a hindrance to the accumulation of those.
What's with the holier-than-thou attitude? Why do you think you're better than someone using chatgpt?
If an AI company has done unethical things do you think it is inappropriate to discuss that? Take Grok: among other things it created sexualized images of underaged women without their consent, not by accident but as a feature. Is that just something you want to ignore? In response the people in charge merely restricted the feature to paid subscribers instead of removing it.
Do you think people who mention grok creating CSAM is a holier-than-thou attitude? Do you not think the people who ignore that are worse than other people?
I assume in any sort of thread on a topic like this there is going to be inorganic activity. These companies are all fighting rather hard to try to gain marketshare, potentially worth $trillions, with a product fully capable of producing endless reasonably compelling content to populate an account, a website, or any other basic proof of identity one might ever want.
It's probably never been the case that plurality of views meant anything since online is a bubble to begin with, filtered by endless biases wherever we happen to be reading, making it an even more fringe bubble, but the advent of AI has pushed it all over the edge to the point that perceived pluralities are just completely and utterly meaningless. Somewhat depressing for a one who enjoys online chat as a pasttime, but it's the reality of the world now.
Yeah yeah yeah, everyone's a bot expect you with all the right opinions...
hehe :)
What's a good way to think about this? Because it does cross my mind about the billions of dollars at play - at the same time, I'm not a pessimist. I think my middle ground is kind of just the usual, taking things with a grain of salt. I mean, I chose to reply to this comment in good faith it's human to human.
I hope I keep that faith. I hope our billions of neighbors on the web enable me to keep that faith over the coming years. Definitely uncertain about the future of the web but want to love it like I've loved it 1990s-today. (Guess I should volunteer w/the EFF while job hunting, try for for-purpose jobs...)
5.4, in my own testing, was almost always ahead of Opus 4.6 for reviews and planning. I'm on plus plan on openai, so I couldn't test it so deeply. Anyone who had more experience on both could perhaps chime in? Pros/cons compared to Opus? I'm invested in Claude ecosystem but the recent quality and session limits decrease have me on the edge.
Same for me. I'm on $20 plan for both and I use them both interchangeably. Similar "intelligence" imo. Just different way of doing things, that's all. But Claude is getting worse in terms of token usage so I've cancelled my plan last month.
Yeah it's probably a bit better overall. 5.4 is a month newer than Opus 4.6
My guess is that 5.5 will come out soon and be significantly better so you'd want to be using Codex then, but then when Opus 5 comes out probably back to claude code
Also 5.4 has fast mode, and higher usage limits since it's cheaper
Do you mind elaborating on your experience here?
Just curious as I've often heard that Claude was superior for planning/architecture work while ChatGPT was superior for actual implementation and finding bugs.
Claude makes more detailed plans that seem better if you just skim them, but when analyzed, has a lot of errors, usually.
It compensates for most during implementation if you make it use TDD by using superpower et al, or just telling it to do so.
GPT 5.4 makes more simple plans (compared to superpowers - a plugin from the official claude plugin marketplace - not the plan mode), but can better fill the details while implementing.
Plan mode in Claude Code got much better in the last months, but the lacking details cannot be compensated by the model during the implementation.
So my workflow has been:
Make claude plan with superpowers:brainstorm, review the spec, make updates, give the spec to gpt, usually to witness grave errors found by gpt, spec gets updates, another manual review, (many iterations later), final spec is written, write the plan, gpt finds mind boggling errors, (many iterations later), claude agent swarm implements, gpt finds even more errors, I find errors, fix fix fix, manual code review and red tests from me, tests get fixed (many iterations later) finally something usable with stylistic issues at most (human opinion)!
This happens with the most complex features that'd be a nightmare to implement even for the most experienced programmers of course. For basic things, most SOTA modals can one-shot anyway.
The title is misleading. The only thing they seem to have done was add a $100 plan identical to Claude's, which gives 5x usage of ChatGPT Plus. There is still a $200 plan that gives 20x usage.
That is not the "only" thing: You get access to GPT-5.4 pro.
Just to clarify, one does not get access to the pro model on the Pro plan?
The $20 Plus plan still exists, and does not give access to the pro model.
The $200 Pro plan still exists, and does give access to the pro model.
What is new is a $100 Pro plan that does give access to the pro model, with lower usage limits than the $200 Pro plan.
This is still worse than Anthropic's right? Because you get access to their top model even at the $20 price point
It's not worse, Anthropic simply has no equivalent model (if you don't consider Mythos) of GPT 5.4 Pro. Google does though: Gemini 3.1 Deep Think.
GPT 5.4 Pro is extremely slow but thorough, so it's not meant for the usual agentic work, rather for research or solving hard bugs/math problems when you provide it all the context.
So, reading the tea leaves, they're either losing subscribers for the $200 plan, or they're not following the same hockey stick path of growth they thought they were... maybe?
Edit: I wonder if this is actually compute-bound as the impetus
Nope, it's just that a lot of people (especially those using Codex) asked us for a medium-sized $100 plan. $20 felt too restrictive and $200 felt like a big jump.
Pricing strategy is always a bit of an art, without a perfect optimum for everyone:
- pay-per-token makes every query feel stressful
- a single plan overcharges light users and annoyingly blocks heavy users
- a zillion plans are confusing / annoying to navigate and change
This change mostly just adds a medium-sized plan for people doing medium-sized amounts of work. People were asking for this, and we're happy to deliver.
(I work at OpenAI.)
Plenty of people wanted to spend more than $20 but less than $200 for a plan. It's long overdue IMO.
Plus plan doesn't get the pro model, which is (AFAICT) the same 5.4 model but thinks like a lot.
You're trying to make words mean what we all think they mean. Stop foisting your Textualism upon us!
Does GPT-5.4 pro give a much better result in some circumstances? What're their typical uses in your experience?
If you want it to deeply research something pro is great. I had a problem I just couldn’t find with my oven so I gave it a lot of information and it went off on its own for about 2 hours and then gave me what I needed to fix the problem (fan was turning off too quickly which was causing the panel to overheat). I have no idea how it figured it out and I couldn’t find anything after hours of googling so it was very impressive. I even went and googled for it once I knew what the problem was and I still couldn’t find the solution that it came up with.
Thanks for sharing this experience. Does it cost a lot of token in the deep analysis - which will make the $100 plan much quicker to drain all budgets.
I think it’s gonna be very hard to blow through your tokens just using chat. I mostly bought the plan so I could use Codex and on the $200 a month plan I’ve basically been using it 15 hours a day almost nonstop and I don’t run out of tokens for the week.
Notably, up until now Pro had 6x usage of Plus. So the title is only slightly misleading.
On the other hand, the benchmark of Plus usage seems to be to be all over the place, so it’s difficult to say now how does the usage compare to the old Pro.
You’re right. I missed the “From $100”. Edited title.
Oh. Yikes.
That has me quite tempted. In general, I stay under the Plus limits, but I do watch my consumption. I could use `/fast` mode all of the time, with extra high reasoning, and use gpt-5.4-pro for especially complex tasks. It wasn't worth 10x the price to me before, but 5x is approachable.
I think you currently can't use pro inside codex, or can you?
I like that they kept limited access to Codex even on free tier.
LE: Someone said this is how the tiers are now counted:
"Essentially if old plus is 1x then new limits are: Plus - 0.3x Pro $100 - 1.5x Pro $200 - 6x (unchanged)"
Are you allowed to run your own autonomous agents with it outside of Codex, like OpenClaw and others?
The era of subsidization is over, it seems.
For my money, on the code side at least, GitHub Copilot on VSCode is still the most cost effective option, 10 bucks for 300 requests gets me all I need, especially when I use OpenAI models which are counted as 1x vs Opus which is 3x. I've stopped using all other tools like Claude Code etc.
I use both GH Copilot as well as CC extensively and it does seem more economical, though I wonder how long this will last as I imagine Github has also been subsidizing LLM usage extensively.
FWIW it feels like GH Copilot is a cheaper version of OpenRouter but with trade-offs like being locked into VSCode and the Microsoft ecosystem overall. I already use VSCode though and otherwise I don't see much downside to using GH Copilot outside of that.
You’re not locked into vscode. There are plugins for other IDEs, and a ‘copilot’ cli tool very similar to Claude Code’s cli tool.
I also wouldn’t say you’re locked into Microsoft’s ecosystem. At work we just have skills that allow for interaction with Bitbucket and other internal tooling. You’re not forced to use GitHub at all.
You can use copilot models from OpenCode:
https://github.blog/changelog/2026-01-16-github-copilot-now-...
I'm hopeful because Microsoft already has a partnership and owns much of OpenAI so can get their models at cost to host on Azure with they already do, so they can pass on the savings to the user. This is why Opus is 3x as expensive in Copilot, because Microsoft needs to buy API usage from Anthropic directly.
I don’t think it’s API costs. Their Sonnet 4.6 is just 1x premium request which matches the 1x cost of the various GPT Codex models.
Sonnet is the worse model though, therefore it's expected that it is cheaper, the comparison would be Opus and GPT. That Anthropic's worse model is the same request cost as the best OpenAI model is what I mean when talking about Microsoft flexing their partnership.
> but with trade-offs like being locked into VSCode and the Microsoft ecosystem overall
You can use GH Copilot with most of Jetbrains IDEs.
You could use something like [https://opencode.ai](OpenCode) which supports integration with Copilot.
OT:
In general I view VS Code and VS.NET Community + SQL Server free universe as the most effective option :) I think these products are great actually.
I tried Claude Code for a week straight recently to see what all the hype was about and while it pumped out a bunch of reasonable looking code and features I ended up feeling completely disconnected from my codebase and uncomfortable.
Cancelled the plan I had with them and happily went back to just coding like normal in VSCode with occasional dips into Copilot when a need arose or for rubber ducking and planning. Feels much better as I'm in full control and not trusting the magic black box to get it right or getting fatigue from reading thousands of lines of generated code.
Anyone who says they're able to review thousands of lines effectively that Claude might slop out in a day are lying to themselves.
I don't like calling a posture 'ignorant' , but I think that's what we have here. I don't mean that as an insult.
It's likely you didn't learn how to use the tool properly, and I'd suggest 'trying again' because not using AI soon will be tantamount to digging holes with shovels instead of using construction equipment.
Yes - AI can generate slop, and probably too many Engineers do that.
Yes - you can 'feel a loss of control' but that's where you have to find your comfort zone.
It's generally a bad idea to produce 'huge amounts of code' - unless it's perfectly consistent with a design, and he architecture is derived from well-known conventions.
Start by using it as an 'assistant' aka research, fill in all the extra bits, and get your testing going.
You'll probably want to guide the architecture, and at least keep an eye on the test code.
Then it's a matter of how much further 'up' you can go,
There are few situations in which we should be 'accepting' large amounts of code, but some of it can be reviewed quickly.
The AI, already now in 2026 can write better code than you at the algorithmic level - it will be tight, clean, 'by the book' and far lesss likley to have erros.
It fails at the architectural and modular level still, that will probably change.
The AI 'makes a clean cut' in the wood, tighter to the line than any carpenter could - like a power tool.
A carpenter that does not use power tools is an 'artisnal craft person' , not really building functional things.
This is the era of motor cars, there is really no option - I don't say that because I'm pro or anti anything, AI is often way over-hyped - that's something else entirely.
It's like the web / cloud etc. it's just 'imminent'.
So try again, experiment, stay open minded.
>Anyone who says they're able to review thousands of lines effectively that Claude might slop out in a day are lying to themselves.
The amount you can review before burning out is now the reasonable limit, for the same reason that a car is supposed to stay at the speed you can handle and not the max speed of the engine.
Of course, many people are secretly skipping reviews and some dare to publicly advocate for getting rid of them entirely.
> For the same reason that a car is supposed to stay at the speed you can handle and not the max speed of the engine.
As we know with driving, sensible drivers stick to the speed limit most of the time, but there's a good percentage of knuckle draggers who just love speeding, some people get drunk, some they just drive the wrong way down the highway entirely. Either way it's usually the sensible people who end up suffering.
why not just use it to review your codebase/commits/prs? you don't have to let it write a bunch of code for you neccessarily.
That's my point - it's great as a tool to talk something through or rubber duck it, but as soon as you just let it loose to slop out thousands of lines a day and never read them all you're really doing is filling your base with thousands of lines of technical debt.
I think most people doing what is now called agentic development, aren't following most established dev methodologies and are to a great extent playing it by vibe.
The codebase disconnect is real.
We are like blue collar workers that need to hit the gym to maintain the body that our cavemen ancestors could maintain by doing their daily duties.
Codebase gym sessions might become a thing.
> The era of subsidization is over
Of course it is. Returns are diminishing, AGI isn't happening with current techniques but it is good enough to sell, so it's time to monetize. I just got an email from OpenAI as well about ads in their free tier (I signed up once out of curiosity).
> AGI isn't happening with current techniques but it is good enough to sell, so it's time to monetize.
Or perhaps it was a scam in the first place for an IPO.
Not over yet. More hikes will come. It will reach $1000.
That's what I said by subsidization being over.
Can you expand what you mean by "subsidization being over" in terms of the plan prices?
- Plus is still the same $20
- 20x Pro is still the same $200
- This is a new 5x tier is $100
https://help.openai.com/en/articles/9793128-what-is-chatgpt-... is probably a better direct comparison of the 3
Codex had 2x usage until April 1, I think when that ended there were a lot more people (like myself) who were fine on $20 but now want more usage
They're trying to slowly move up market. I assume soon the $20 will get its own restrictions in the future (and/or ads) to get people to pay $100.
The same could have been claimed in 2024 when they introduced the $200 Pro plan. Nothing is over yet just because of what could possibly happen next.
Yes, and now they (will) have ads in the lower tiers and eventually will add them in higher tiers as well. It's no different to Netflix's methods. Therefore the statement would have been proven right if it were claimed back then.
But these are, again, additional tiers.
You can't just say because they've added more things the old things are over - the old things actually have to go away first. Eventually they may get there (or not). It may be another few years (or not). Nothing is actually now over though any more than it was now over in 2024.
Additional tiers to get people to move up as they enshittify the lower tiers. We already see it in other companies as well as OpenAI themselves so my inference is based on that, not to wait and see until they do indeed enshittify it.
The subsidization being "over" would mean we are paying their actual cost or more.
I see what you mean, I misread your initial reply.
Pro used to be $200.
Any idea way "5x or 20x more usage" means?
What’s the difference between the two Pro plans?
Both Pro plans include the same core capabilities. The main difference is usage allowance: Pro $100 unlocks 5x higher usage than Plus (and 10x Codex usage vs. Plus for a limited time), while Pro $200 unlocks 20x usage than Plus.
From their faq
5x more usage than in Plus is 100$
20x more usage than in Plus is 200$
I see this when I try to upgrade my Plus subscription.
If you pay 200 you get 20x
The price is $100 according to this post. Where is there an option for $200?
You choose on checkout. There it says
So curious that the cost in the comparison is just a flat $100, not "$100 or $200" and yet the usage has the "or". Surely just a lapse in copy editing.
Surely they weren't trying to be deceptive... surely.
Anthropic is the exact same way, I think they're just trying to avoid having 5 different subscription tiers visible. Probably needing 20x is very niche
It states “From $100”. Standard pricing speak.
seems like this $100 replaced the $200 plan
So.. cheaper?
No, the same $200 plan is still there. They hid it behind the $100 click-through.
This just adds a $100 plan that's 1/4 the usage of the $200 plan..
I assume it means 5x if they get to choose. They're the ones enforcing the limits.
Suppose you enter a casino and the owner welcomes you in and sees that you are a frequent loyal s̶p̶e̶n̶d̶e̶r̶ customer (with the amount of tokens you are spending a month) with an existing membership.
With this new VIP membership that comes with 5x or 20x usage, if you spend $100 you get 5x. $200 you get 20x and you get to spin the wheel and use the slot machines unlimited times even at peak hours more than most without any restrictions, 24/7, no waiting for hours with priority.
So spend more to get more abundance and more simultaneous spins at the wheel.
Except if you're trying to abuse the slot machines themselves or sharing or reselling your membership to other customers who want a spin at the roulette wheel; but were previously banned. [0]
[0] https://help.openai.com/en/articles/9793128-about-chatgpt-pr...
Exactly! :)
They are actively exploiting the compute shortages of Anthropic. In our team we're pushing for more or less vanilla and portability, since the best harness today might not be the best one in 6 months.
This is an additional offering to the existing plan.
5x=$100 20x=$200
It looks like its called prolite.
https://snipboard.io/jmGKfM.jpg
How much was it before?
$20 or $200 plan, now we have a middle plan.
Does this give you something different than the $20/mo plan when using codex?
Yes, it's 5x more usage than Plus, and with the current promotions you actually get 10x more usage than Plus on the $100 plan until May 31st.
Same for the $200 plan, it's still 2x its normal usage until that date.
For me it's not the price. It's the fact that they obviously read my prompts and may even use a derived version of my data for training. As it's very clear in the meantime that SAMA lies most of the time, there's just no way I can trust this company in any way.
are your prompts that important that you would not use SOTA models just to protect them?
For me, they are just a means to an end and disposable.
So.. would you mind releasing all your code on GitHub?
https://x.com/OpenAI/status/2042296046009626989
>Our existing $200 Pro tier still remains our highest usage option.
Awesome news.
And that includes usage of the API with any agent without risking being banned. OpenAI is also very supportive of open source software.
I'm using GPT-5.4 with Swival (https://swival.dev) for a while, alongside local models, and it's absolutely fantastic.
This is like the 2010s hosting price wars.
What is the difference between Pro and normal mode apart from the fact the Pro takes ages to finish? I see not much difference in output quality.
It really feels like LLMs will mostly become tools for tech workers rather than the kind of civilization-level transformation sama has been peddling. Every single comment here seems to confirm the above.
The missing piece isn't smarter models, it's making them usable for normal people. OpenClaw proved there's massive demand for AI agents but the setup and maintenance is brutal. We built Atmita (https://atmita.com) from scratch, not based on OpenClaw, cloud-native so there's nothing to install. It can hit any REST API directly, integrates with 800+ services out of the box, has a built-in browser, runs scheduled tasks in the background, and has built-in approval controls. All through a chat interface anyone can use.
This is kind of the goldilocks zone for LLMs right now.
I wouldn't mistake this for any kind of capability plateau. There is a massive push towards making transformers the engine of humanoid (and other kinds of) robotics, we just haven't reached the hype moment for those yet.
> I wouldn't mistake this for any kind of capability plateau. There is a massive push towards making transformers the engine of humanoid (and other kinds of) robotics, we just haven't reached the hype moment for those yet.
Problem is that the fuel to get this train going relies on investors money. Investors aren't going to be happy with the quote I took from your message.
And that's the real bet really, can the industry turn the spark into fire before the investor money runs out?
I have heard a CTO had a major success building a side project over a weekend.
Is that CTO in the room with us?
Tell me you're losing market share to competitors without telling me you're losing market share to competitors
Can you guys remind me again why you're doing this?
just a rumor, but i heard altman was adding a timer which required the R&D dept. to triple
I heard it’ll take about a year. Timers are a hard problem to solve.
While they work on getting it to tell the time they'll just be over there listing targets for military strikes.
Price drops are nice. Unfortunately, the quality differential versus the competitor is night and day.
And everyone serious uses the API rate billing anyway.
> the quality differential versus the competitor is night and day.
This myth about the inferiority of ChatGPT and Codex is becoming a meme.
I have active subscriptions to both. I am throwing at Codex all kinds of data engineering, web development and machine learning problems, have been working on non-tech tasks in the "Karpathy Obsidian Wiki" [1] style before he posted about it.
Not only does Codex crush Claude on cost, it's also significantly better at adherence and overall quality. Claude is there on my Mac, gathering dust, to the point I am thinking of not renewing the sub.
There are plenty of fellow HNers here who feel the same from what I read in the flamewars. I suspect none of us really has a horse in this race and many are half-competent (in other threads, they mention they do things like embedded programming, distributed DL systems, etc.)
I'm starting to suspect a vast majority of people pushing the narrative that Claude is vastly better haven't even tried the 5.3 / 5.4 models and are doing it out of sheer tribalism.
[1] https://gist.github.com/karpathy/442a6bf555914893e9891c11519...
I have access to effectively infinite API tokens for all models from Anthropic as well as OpenAI. The differential in performance in complex tasks is vast and strongly in favor of Opus, in my experience. I do not use the official harnesses for either model, though - as they are not my taste.
Codex is closer to my taste, as it is at least a native app and not typescript slop. But the model is just not up to snuff.
Disagree. I use codex extensively. It just works so well with vscode and python. Claude with ridiculous limits - thanks no. For some even xAI is good fit.
This take is out-of-date by months (which is an eternity in this space). Codex today has caught up and is very much on par with CC.
I prefer and use 5.4 over Opus, it's simply better, faster, and doesn't glaze me like Claude models want to do for some reason.