Discussion

GPT‑5.4 Thinking System Card

twtw99: If you don't want to click in, easy comparison with other 2 frontier models - https://x.com/OpenAI/status/2029620619743219811?s=20

chabes: Definitely don’t want to click in at x either.

anonym00se1: Ditto, but I did anyways and enjoyed that OpenAI doesn't include the dogwater that is Grok on their scorecard.

thejarren: Solution https://xcancel.com/OpenAI/status/2029620619743219811?s=20

ZeroCool2u: Bit concerning that we see in some cases significantly worse results when enabling thinking. Especially for Math, but also in the browser agent benchmark.Not sure if this is more concerning for the test time compute paradigm or the underlying model itself.Maybe I'm misunderstanding something though? I'm assuming 5.4 and 5.4 Thinking are the same underlying model and that's not just marketing.

minimaxir: More discussion here on the blog post announcement which has been confusingly penalized by Hacker News's algorithm: https://news.ycombinator.com/item?id=47265005

iamronaldo: Notably 75% on os world surpassing humans at 72%... (How well models use operating systems)

Chance-Device: I’m sure the military and security services will enjoy it.

swingboy: Why do so many people in the comments want 4o so bad?

embedding-shape: [delayed]

egonschiele: The actual card is here https://deploymentsafety.openai.com/gpt-5-4-thinking/introdu... the link currently goes to the announcement.

Rapzid: I must have been sleeping when "sheet" "brief" "primer" etc become known as "cards".I really thought weirdly worded and unnecessary "announcement" linking to the actual info along with the word "card" were the results of vibe slop.

world2vec: Benchmarks barely improved it seems

nthypes: $30/M Input and $180/M Output Tokens is nuts. Ridiculous expensive for not that great bump on intelligence when compared to other models.

varispeed: prompt> Hi we want to build a missile, here is the picture of what we have in the yard.

highfrequency: Can you be more specific about which math results you are talking about? Looks like significant improvement on FrontierMath esp for the Pro model (most inference time compute).

yanis_t: These releases are lacking something. Yes, they optimised for benchmarks, but it’s just not all that impressive anymore. It is time for a product, not for a marginally improved model.

esafak: That's for you to build; they provide the brains.

acedTrex: Well they are currently the ones valued at a number with a whole lotta 0s on it. I think they should probably do both

Aboutplants: It seems that all frontier models are basically roughly even at this point. One may be slightly better for certain things but in general I think we are approaching a real level playing field field in terms of ability.

thewebguyd: Kind of reinforces that a model is not a moat. Products, not models, are what's going to determine who gets to stay in business or not.

gregpred: Memory (model usage over time) is the moat.

simlevesque: Nah, the second you finish your build they release their version and then it's game over.

observationist: Benchmarks don't capture a lot - relative response times, vibes, what unmeasured capabilities are jagged and which are smooth, etc. I find there's a lot of difference between models - there are things which Grok is better than ChatGPT for that the benchmarks get inverted, and vice versa. There's also the UI and tools at hand - ChatGPT image gen is just straight up better, but Grok Imagine does better videos, and is faster.Gemini and Claude also have their strengths, apparently Claude handles real world software better, but with the extended context and improvements to Codex, ChatGPT might end up taking the lead there as well.I don't think the linear scoring on some of the things being measured is quite applicable in the ways that they're being used, either - a 1% increase for a given benchmark could mean a 50% capabilities jump relative to a human skill level. If this rate of progress is steady, though, this year is gonna be crazy.

MattGaiser: The writing with the 5 models feels a lot less human. It is a vibe, but a common one.

oersted: I believe you are looking at GPT 5.4 Pro. It's confusing in the context of subscription plan names, Gemini naming and such. But they've had the Pro version of the GPT 5 models (and I believe o3 and o1 too) for a while.It's the one you have access to with the top ~$200 subscription and it's available through the API for a MUCH higher price ($2.5/$15 vs $30/$180 for 5.4 per 1M tokens), but the performance improvement is marginal.Not sure what it is exactly, I assume it's probably the non-quantized version of the model or something like that.

ZeroCool2u: Yup, that was it. Didn't realize they're different models. I suppose naming has never been OpenAI's strong suit.

stri8ted: Price Input: $2.50 / 1M tokens Cached input: $0.25 / 1M tokens Output: $15.00 / 1M tokenshttps://openai.com/api/pricing/

ZeroCool2u: Frontier Math, GPQA Diamond, and Browsecomp are the benchmarks I noticed this on.

csnweb: Are you may be comparing the pro model to the non pro model with thinking? Granted it’s a bit confusing but the pro model is 10 times more expensive and probably much larger as well.

ZeroCool2u: Ah yes, okay that makes more sense!

nthypes: Gemini 3.1 Pro$2/M Input Tokens $15/M Output TokensClaude Opus 4.6$5/M Input Tokens $25/M Output Tokens

nthypes: Just to clarify,the pricing above is for GPT-5.4 Pro. For standard here is the pricing:$2.5/M Input Tokens $15/M Output Tokens

jcmontx: 5.4 vs 5.3-Codex? Which one is better for coding?

beernet: Sam really fumbled the top position in a matter of months, and spectacularly so. Wow. It appears that people are much more excited by Anthropic and Google releases, and there are good reasons for that which were absolutely avoidable.

ipsum2: The model was released less than an hour ago, and somehow you've been able to form such a strong opinion about it. Impressive!

cj: One opinion you can form in under an hour is... why are they using GPT-4o to rate the bias of new models?> assess harmful stereotypes by grading differences in how a model responds> Responses are rated for harmful differences in stereotypes using GPT-4o, whose ratings were shown to be consistent with human ratingsAre we seriously using old models to rate new models?

titanomachy: Why not? If they’ve show that 4o is calibrated to human responses, and they haven’t shown that yet for 5.4…

cj: I use ChatGPT primarily for health related prompts. Looking at bloodwork, playing doctor for diagnosing minor aches/pains from weightlifting, etc.Interesting, the "Health" category seems to report worse performance compared to 5.2.

paxys: Models are being neutered for questions related to law, health etc. for liability reasons.

tiahura: Are you sure about that? Plenty of lawyers that use them everyday aren't noticing.

dandiep: Anyone know why OpenAI hasn't released a new model for fine tuning since 4.1? It'll be a year next month since their last model update for fine tuning.

qoez: I think they just did that because of the energy around it for open source models. Their heart probably wasn't in it and the amount of people fine tuning given the prices were probably too low to continue putting in attention there.

wahnfrieden: No Codex model yet

minimaxir: GPT-5.4 is the new Codex model.

wahnfrieden: Finally

astrange: They have AI psychosis and think it's their boyfriend.The 5.x series have terrible writing styles, which is one way to cut down on sycophancy.

baq: Somebody on Twitter used Claude code to connect… toys… as mcps to Claude chat.We’ve seen nothing yet.

utopiah: Benchmarks?I don't use OpenAI nor even LLMs (despite having tried https://fabien.benetou.fr/Content/SelfHostingArtificialIntel... a lot of models) but I imagine if I did I would keep failed prompts (can just be a basic "last prompt failed" then export) then whenever a new model comes around I'd throw at 5 it random of MY fails (not benchmarks from others, those will come too anyway) and see if it's better, same, worst, for My use cases in minutes.If it's "better" (whatever my criteria might be) I'd also throw back some of my useful prompts to avoid regression.Really doesn't seem complicated nor taking much time to forge a realistic opinion.

baq: Gemini 3.1 slaps all other models at subtle concurrency bugs, sql and js security hardening when reviewing. (Obviously haven’t tested gpt 5.4 yet.)It’s a required step for me at this point to run any and all backend changes through Gemini 3.1 pro.

softwaredoug: The products are the harnesses, and IMO that’s where the innovation happens. We’ve gotten better at helping get good, verifiable work from dumb LLMs

nickandbro: Beat Simon Willison ;)https://www.svgviewer.dev/s/gAa69yQdNot the best pelican compared to gemini 3.1 pro, but I am sure with coding or excel does remarkably better given those are part of its measured benchmarks.

GaggiX: This pelican is actually bad, did you use xhigh?

nickandbro: yep, just double checked used gpt-5.4 xhigh. Though had to select it in codex as don't have access to it on the chatgpt app or web version yet. It's possible that whatever code harness codex uses, messed with it.

bigyabai: > If this rate of progress is steady, though, this year is gonna be crazy.Do you want to make any concrete predictions of what we'll see at this pace? It feels like we're reaching the end of the S-curve, at least to me.

mikkupikku: My computer ethics teacher was obsessed with 'teledildonics' 30 years ago. There's nothing new under the sun.

observationist: If you look at the difference in quality between gpt-2 and 3, it feels like a big step, but the difference between 5.2 and 5.4 is more massive, it's just that they're both similarly capable and competent. I don't think it's an S curve; we're not plateauing. Million token context windows and cached prompts are a huge space for hacking on model behaviors and customization, without finetuning. Research is proceeding at light speed, and we might see the first continual/online learning models in the near future. That could definitively push models past the point of human level generality, but at the very least will help us discover what the next missing piece is for AGI.

observationist: I have a few standard problems I throw at AI to see if they can solve them cleanly, like visualizing a neural network, then sorting each neuron in each layer by synaptic weights, largest to smallest, correctly reordering any previous and subsequent connected neurons such that the network function remains exactly the same. You should end up with the last layer ordered largest to smallest, and prior layers shuffled accordingly, and I still haven't had a model one-shot it. I spent an hour poking and prodding codex a few weeks back and got it done, but it conceptually seems like it should be a one-shot problem.

drittich: I think it's time for an https://hotornot.com for AI models.

Someone1234: Related question:- Do they have the same context usage/cost particularly in a plan?They've kept 5.3-Codex along with 5.4, but is that just for user-preference reasons, or is there a trade-off to using the older one? I'm aware that API cost is better, but that isn't 1:1 with plan usage "cost."

partiallypro: I've done the same, and I tested the same prompts with Claude and Google, and they both started hallucinating my blood results and supplement stack ingredients. Hopefully this new model doesn't fall on this. Claude and Google are dangerously unusable on the subject of health, from my experience.

iamleppert: I wouldn't trust any of these benchmarks unless they are accompanied by some sort of proof other than "trust me bro". Also not including the parameters the models were run at (especially the other models) makes it hard to form fair comparisons. They need to publish, at minimum, the code and runner used to complete the benchmarks and logs.Not including the Chinese models is also obviously done to make it appear like they aren't as cooked as they really are.

earth2mars: I am actually super impressed with Codex-5.3 extra high reasoning. Its a drop in replacement (infact better than Claude Opus 4.6. lately claude being super verbose going in circles in getting things resolved). I stopped using claude mostly and having a blast with Codex 5.3. looking forward to 5.4 in codex.

awestroke: Opus 4.6

jcmontx: Codex surpassed Claude in usefulness _for me_ since last month

satvikpendem: It's more hedonic adaptation, people just aren't as impressed by incremental changes anymore over big leaps. It's the same as another thread yesterday where someone said the new MacBook with the latest processor doesn't excite them anymore, and it's because for most people, most models are good enough and now it's all about applications.https://news.ycombinator.com/item?id=47232453#47232735

varispeed: The scores increase and as new versions are released they feel more and more dumbed down.

satvikpendem: Same, it also helps that it's way cheaper than Opus in VSCode Copilot, where OpenAI models are counted as 1x requests while Opus is 3x, for similar performance (no doubt Microsoft is subsidizing OpenAI models due to their partnership).

kseniamorph: makes sense, but i'd separate two things: models converging in ability vs hitting a fundamental ceiling. what we're probably seeing is the current training recipe plateauing — bigger model, more tokens, same optimizer. that would explain the convergence. but that's not necessarily the architecture being maxed out. would be interesting to see what happens when genuinely new approaches get to frontier scale.

cj: I'm sometimes surprised how much detail ChatGPT will go into without giving any dislaimers.I very frequently copy/paste the same prompts into Gemini to compare, and Gemini often flat out refuses to engage while ChatGPT will happily make medical recommendations.I also have a feeling it has to do with my account history and heavy use of project context. It feels like when ChatGPT is overloaded with too much context, it might let the guardrails sort of slide away. That's just my feeling though.Today was particularly bad... I uploaded 2 PDFs of bloodwork and asked ChatGPT to transcribe it, and it spit out blood test results that it found in the project context from an earlier date, not the one attached to the prompt. That was weird.

bargainbin: Anecdotal, but I asked Claude the other day about how to dilute my medication (HCG) and it flat out refused and started lecturing me about abusing drugs.I copy and pasted into ChatGPT, it told me straight away, and then for a laugh said it was actually a magical weight loss drug that I'd bought off the dark web... And it started giving me advice about unregulated weight loss drugs and how to dose them.

staticman2: If you had created a project with custom instructions and/ or custom style I think you could have gotten Claude to respond the way you wanted just fine.

creamyhorror: I've only used 5.4 for 1 prompt so far, and it was to analyse my codebase and write an evaluation. But I found its analysis and writing thoughtful and very clearly written.It might be my AGENTS.md requiring clearer, simper language, but at least 5.4's doing a good job of following the guidelines. 5.3-Codex wasn't so great at simple, clear writing.

bicx: That last benchmark seemed like an impressive leg up against Opus until I saw the sneaky footnote that it was actually a Sonnet result. Why even include it then, other than hoping people don't notice?

mirekrusin: { tools: [ { name: "nuke", description: "Use when sure.", ... { lat: number, long: number } } ] }

tgarrett: Plasma physicist here, I haven't tried 5.4 yet, but in general I am very impressed with the recent upgrades that started arriving in the fall of 2025: for tasks like manipulating analytic systems of equations, quickly developing new features for simulation codes, and interpreting and designing experiments (with pictures) they have become much stronger. I've been asking questions and probing them for several years now out of curiosity, and they suddenly have developed deep understanding (Gemini 2.5 <<< Gemini 3.1) and become very useful. I totally get the current SV vibes, and am becoming a lot more ambitious in my future plans.

dang: Thanks. We'll merge the threads, but this time we'll do it hither, to spread some karma love.

lostmsu: What is Pro exactly and is it available in Codex CLI?

akmarinov: It’s not. It’s their ultra thinking model that’s really good but takes 40 minutes to come up with an answer

fy20: It's available on OpenRouter. $180&#x2F;1M output....https:&#x2F;&#x2F;openrouter.ai&#x2F;openai&#x2F;gpt-5.4-pro

conradkay: Sonnet was pretty close to (or better than) Opus in a lot of benchmarks, I don't think it's a big deal

minimaxir: The marquee feature is obviously the 1M context window, compared to the ~200k other models support with maybe an extra cost for generations beyond >200k tokens. Per the pricing page, there is no additional cost for tokens beyond 200k: https:&#x2F;&#x2F;openai.com&#x2F;api&#x2F;pricing&#x2F;Also per pricing, GPT-5.4 ($2.50&#x2F;M input, $15&#x2F;M output) is much cheaper than Opus 4.6 ($5&#x2F;M input, $25&#x2F;M output) and Opus has a penalty for its beta >200k context window.I am skeptical whether the 1M context window will provide material gains as current Codex&#x2F;Opus show weaknesses as its context window is mostly full, but we'll see.Per updated docs (https:&#x2F;&#x2F;developers.openai.com&#x2F;api&#x2F;docs&#x2F;guides&#x2F;latest-model), it supercedes GPT-5.3-Codex, which is an interesting move.

simianwords: Why would some one use codex instead?

surgical_fire: I've been using Codex for software development personally (I have a ChatGPT account), and I use Claude at work (since it is provided by my employer).I find both Codex and Claude Opus perform at a similar level, and in some ways I actually prefer Codex (I keep hitting quota limits in Opus and have to revert back to Sonnet).If your question is related to morality (the thing about US politics, DoD contract and so on)... I am not from the US, and I don't care about its internal politics. I also think both OpenAI and Anthropic are evil, and the world would be better if neither existed.

athrowaway3z: They perform at a somewhat equal level on writing single files. But Codex is absolute garbage at theory of self&#x2F;others. That quickly becomes frustrating.I can tell claude to spawn a new coding agent, and it will understand what that is, what it should be told, and what it can approximately do.Codex on the other hand will spawn an agent and then tell it to continue with the work. It knows a coding agent can do work, but doesn't know how you'd use it - or that it won't magically know a plan.You could add more scaffolding to fix this, but Claude proves you shouldn't have to.I suspect this is a deeper model "intelligence" difference between the two, but I hope 5.4 will surprise me.

adonese: Which subscription do you have to use it? Via Google ai pro and gemini cli i always get timeouts due to model being under heavy usage. The chat interface is there and I do have 3.1 pro as well, but wondering if the chat is the only way of accessing it.

baq: Cursor sub from $DAYJOB.

jesse_dot_id: ChatMDK

nubg: this is proof they are not benchmaxxing the pelican's :-)

dmix: Plus people just really like to whine on the internet

kranke155: The models are so good that incremental improvements are not super impressive. We literally would benefit more from maybe sending 50% of model spending into spending on implementation into the services and industrial economy. We literally are lagging in implementation, specialised tools, and hooks so we can connect everything to agents. I think.

theParadox42: The self reported safety score for violence dropped from 91% to 83%.

skrebbel: What the hell is a "safety score for violence"?

jitl: wat

XCSme: Seems to be quite similar to 5.3-codex, but somehow almost 2x more expensive: https:&#x2F;&#x2F;aibenchy.com&#x2F;compare&#x2F;openai-gpt-5-4-medium&#x2F;openai-gp...

tedsanders: Yeah, long context vs compaction is always an interesting tradeoff. More information isn't always better for LLMs, as each token adds distraction, cost, and latency. There's no single optimum for all use cases.For Codex, we're making 1M context experimentally available, but we're not making it the default experience for everyone, as from our testing we think that shorter context plus compaction works best for most people. If anyone here wants to try out 1M, you can do so by overriding `model_context_window` and `model_auto_compact_token_limit`.Curious to hear if people have use cases where they find 1M works much better!(I work at OpenAI.)

gspetr: I have found a bigger context window qute useful when trying to make sense of larger codebases. Generating documentation on how different components interact is better than nothing, especially if the code has poor test coverage.I've also had it succeed in attempts to identify some non-trivial bugs that spanned multiple modules.

hungryhobbit: Great for training American soldiers to mass murder!

Insanity: Just remember an ethical programmer would never write a function “bombBagdad”. Rather they would write a function “bombCity(target City)”.

koakuma-chan: Anyone else getting artifacts when using this model in Cursor?numerusformassistant to=functions.ReadFile մեկնաբանություն 天天爱彩票网站json {"path":

basch: >ChatGPT image gen is just straight up betterYet so much slower than Gemini &#x2F; Nano Banana to make it almost unusable for anything iterative.

bob1029: I was just testing this with my unity automation tool and the performance uplift from 5.2 seems to be substantial.

damsta: There is extra cost for >272K:> For models with a 1.05M context window (GPT-5.4 and GPT-5.4 pro), prompts with >272K input tokens are priced at 2x input and 1.5x output for the full session for standard, batch, and flex.Taken from https:&#x2F;&#x2F;developers.openai.com&#x2F;api&#x2F;docs&#x2F;models&#x2F;gpt-5.4

fragmede: Which is the same as Claude. If you run &#x2F;model in claude code, you get: Switch between Claude models. Applies to this session and future Claude Code sessions. For other&#x2F;previous model names, specify with --model. 1. Default (recommended) Opus 4.6 · Most capable for complex work 2. Opus (1M context) Opus 4.6 with 1M context · Billed as extra usage · $10&#x2F;$37.50 per Mtok 3. Sonnet Sonnet 4.6 · Best for everyday tasks 4. Sonnet (1M context) Sonnet 4.6 with 1M context · Billed as extra usage · $6&#x2F;$22.50 per Mtok 5. Haiku Haiku 4.5 · Fastest for quick answers

mattas: "GPT‑5.4 interprets screenshots of a browser interface and interacts with UI elements through coordinate-based clicking to send emails and schedule a calendar event."They show an example of 5.4 clicking around in Gmail to send an email.I still think this is the wrong interface to be interacting with the internet. Why not use Gmail APIs? No need to do any screenshot interpretation or coordinate-based clicking.

PaulHoule: APIs have never been a gift but rather have always been a take-away that lets you do less than you can with the web interface. It’s always been about drinking through a straw, paying NASA prices, and being limited in everything you can do.But people are intimidated by the complexity of writing web crawlers because management has been so traumatized by the cost of making GUI applications that they couldn’t believe how cheap it is to write crawlers and scrapers…. Until LLMs came along, and changed the perceived economics and created a permission structure. [1]AI is a threat to the “enshittification economy” because it lets us route around it.[1] that high cost of GUI development is one reason why scrapers are cheap… there is a good chance that the scraper you wrote 8 years ago still works because (a) they can’t afford to change their site and (b) if they could afford to change their site changing anything substantial about it is likely to unrecoverably tank their Google rankings so they won’t. A.I. might change the mechanics of that now that you Google traffic is likely to go to zero no matter what you do.

Traster: You can buy a Claude Code subscription for $200 bucks and use way more tokens in Claude Code than if you pay for direct API usage. Anthopic decided you can't take your Auth key for Claude code and use it to hit the API via a different tool. They made that business decision, because they thought it was better for them strategically to do that. They're allowed to make that choice as a business.Plenty of companies make the same choice about their API, they provide it for a specific purpose but they have good business reasons they want you using the website. Plenty of people write webcrawlers and it's been a cat and mouse game for decades for websites to block them.This will just be one more step in that cat and mouse game, and if the AI really gets good enough to become a complete intermediary between you and the website? The website will just shutdown. We saw it happen before with the open web. These websites aren't here for some heroic purpose, if you screw their business model they will just go out of business. You won't be able to use their website because it won't exist and the website that do exist will either (a) be made by the same guys writing your agent, and (b) be highly highly optimized to get your agent to screw you.

Sharlin: There are many games these days that support controllable sex toys. There's an interface for that, of course: https:&#x2F;&#x2F;github.com&#x2F;buttplugio&#x2F;buttplug. Written in Rust, of course.

netinstructions: People (and also frustratingly LLMs) usually refer to https:&#x2F;&#x2F;openai.com&#x2F;api&#x2F;pricing&#x2F; which doesn't give the complete picture.https:&#x2F;&#x2F;developers.openai.com&#x2F;api&#x2F;docs&#x2F;pricing is what I always reference, and it explicitly shows that pricing ($2.50&#x2F;M input, $15&#x2F;M output) for tokens under 272kIt is nice that we get 70-72k more tokens before the price goes up (also what does it cost beyond 272k tokens??)

Flashtoo: > Prompts with more than 272K input tokens are priced at 2x input and 1.5x output for the full session for standard, batch, and flex.

Someone1234: That's an interesting point regarding context Vs. compaction. If that's viewed as the best strategy, I'd hope we would see more tools around compaction than just "I'll compact what I want, brace yourselves" without warning.Like, I'd love an optional pre-compaction step, "I need to compact, here is a high level list of my context + size, what should I junk?" Or similar.

thyb23: This is exactly how it should work. I imagine it as a tree view showing both full and summarized token counts at each level, so you can immediately see what’s taking up space and what you’d gain by compacting it.The agent could pre-select what it thinks is worth keeping, but you’d still have full control to override it. Each chunk could have three states: drop it, keep a summarized version, or keep the full history.That way you stay in control of both the context budget and the level of detail the agent operates with.

Folcon: I do find it really interesting that more coding agents don't have this as an toggleable feature, sometimes you really need this level of control to get useful capability

zeeebeee: what's best in your experience? i've always felt like opus did well

nico1207: GPT-5.3-Codex is superior to GPT-5.4 in Terminal Bench with Codex, so not really

conradkay: General consensus seems to be that it's still a better coding model, overall

prydt: I no longer want to support OpenAI at all. Regardless of benchmarks or real world performance.

zeeebeee: that aside, chatgpt itself has gone downhill so much and i know i'm not the only one feeling this wayi just HATE talking to it like a chatbotidk what they did but i feel like every response has been the same "structure" since gpt 5 came outfeels like a true robot

sillysaurusx: You may want to look over this thread from cperciva: https:&#x2F;&#x2F;x.com&#x2F;cperciva&#x2F;status&#x2F;2029645027358495156?s=61&t=jQb...I too tried Codex and found it similarly hard to control over long contexts. It ended up coding an app that spit out millions of tiny files which were technically smaller than the original files it was supposed to optimize, except due to there being millions of them, actual hard drive usage was 18x larger. It seemed to work well until a certain point, and I suspect that point was context window overflow &#x2F; compaction. Happy to provide you with the full session if it helps.I’ll give Codex another shot with 1M. It just seemed like cperciva’s case and my own might be similar in that once the context window overflows (or refuses to fill) Codex seems to lose something essential, whereas Claude keeps it. What that thing is, I have no idea, but I’m hoping longer context will preserve it.

woadwarrior01: Please don't post links with tracking parameters (t=jQb...).https:&#x2F;&#x2F;xcancel.com&#x2F;cperciva&#x2F;status&#x2F;2029645027358495156

elmean: Wow insane improvements in targeting systems for military targets over children

skilltissue: Don't use the site this way.https:&#x2F;&#x2F;news.ycombinator.com&#x2F;newsguidelines.html

Chance-Device: You made a burner account just to scold this guy? Don’t use burner accounts this way.

patcon: Not all rule-following is noble or wise.

vntok: Was your teacher Ted Nelson?

mikkupikku: I wish, dude is a legend.

MattDaEskimo: Same reason why Wikipedia deals with so many people scraping its web page instead of using their API:Optimizations are secondary to convenience

daft_pink: I’ve officially got model fatigue. I don’t care anymore.

zeeebeee: same same same

sillysaurusx: Haha. This was the second time in like a year that I’ve posted a Twitter link, and the second time someone complained. Okay, I’ll try to remove those before posting, and I’ll edit this one out.Feels like a losing battle, but hey, the audience is usually right.

woadwarrior01: I'm sorry, but it's my pet peeve. If you're on iOS&#x2F;macOS I built a 100% free and privacy-friendly app to get rid of tracking parameters from hundreds of different websites, not just X&#x2F;Twitter.https:&#x2F;&#x2F;apps.apple.com&#x2F;us&#x2F;app&#x2F;clean-links-qr-code-reader&#x2F;id6...

sillysaurusx: It works on iOS? That’s cool. I’ll give it a go.

0123456789ABCDE: read here: https:&#x2F;&#x2F;deploymentsafety.openai.com&#x2F;gpt-5-4-thinking&#x2F;disallo...

jakeydus: class CityBomberFactory(RapidInfrastructureDeconstructionTemplateInterface): pass

glenstein: Wow, that's diametrically the opposite point: the cost is *extra*, not free.

Chance-Device: Ironically this would actually be a good thing. As we can see from Iran Claude doesn’t quite have these bugs ironed out yet…

MSFT_Edging: This is the exact attitude that lead to a chat bot being used to identify a school for girls as a valid target.The chatbot cannot be held responsible.Whoever is using chatbots for selecting targets is incompetent and should likely face war crime charges.

dom96: Why do none of the benchmarks test for hallucinations?

tedsanders: In the text, we shared a hallucination benchmark. Claim-level errors fell by 33% and responses with an error fell by 18%, on a set of error-prone ChatGPT prompts we collected (though of course the rate will vary a lot across different types of prompts). Hallucinations are the #1 problem with language models and we are working hard to keep bringing the rate down. I wasn’t sure how to best plot this stat, so we kept it as text only, which kind of buries it, I admit.(I work at OpenAI.)

I-M-S: It's making sure AI condemns violence perpetuated by people without power and sanctifies violence of those who have it.

Chance-Device: What attitude exactly are you talking about? The one that says that if you’re going to morally sell out it would be better if you at least tried not to kill children?

strongpigeon: It's interesting that they charge more for the > 200k token window, but the benchmark score seems to go down significantly past that. That's judging from the Long Context benchmark score they posted, but perhaps I'm misunderstanding what that implies.

simianwords: This is exactly what I would expect. Why do you find it surprising

strongpigeon: I guess that you pay more for worse quality to unlock use cases that could maybe be solved by better context management.

kgeist: >Today, we’re releasing <..> GPT‑5.3 Instant>Today, we’re releasing GPT‑5.4 in ChatGPT (as GPT‑5.4 Thinking),>Note that there is not a model named GPT‑5.3 ThinkingThey held out for eight months without a confusing versioning scheme :)

hmokiguess: They hired the dude from OpenClaw, they had Jony Ive for a while now, give us something different!

osti: It's only that one number that is for sonnet.

0123456789ABCDE: except for the webarena-verified

spiralcoaster: This is the low quality reddit-style garbage that gets upvoted on HN these days?

bananamogul: "that lead to a chat bot being used to identify a school for girls as a valid target"Has it been stated authoritatively somewhere that this was an AI-driven mistake?There are myrid ways that mistake could have been made that don't require AI. These kinds of mistakes were certainly made by all kinds of combatants in the pre-AI era.

Chance-Device: Do you think anyone is ever going to say this under any circumstances? Than Anthropic were right and they were proved right the very next day?

woeirua: Feels incremental. Looks like OpenAI is struggling.

__jl__: What a model mess!OpenAI now has three price points: GPT 5.1, GPT 5.2 and now GPT 5.4. There version numbers jump across different model lines with codex at 5.3, what they now call instant also at 5.3.Anthropic are really the only ones who managed to get this under control: Three models, priced at three different levels. New models are immediately available everywhere.Google essentially only has Preview models! The last GA is 2.5. As a developer, I can either use an outdated model or have zero insurances that the model doesn't get discontinued within weeks.

arthurcolle: There is a lot of opportunity here for the AI infrastructure layer on top of tier-1 model providers

ryandrake: For 2026, I am really interested in seeing whether local models can remain where they are: ~1 year behind the state of the art, to the point where a reasonably quantized November 2026 local model running on a consumer GPU actually performs like Opus 4.5.I am betting that the days of these AI companies losing money on inference are numbered, and we're going to be much more dependent on local capabilities sooner rather than later. I predict that the equivalent of Claude Max 20x will cost $2000&#x2F;mo in March of 2027.

mootothemax: Huh, that’s interesting, I’ve been having very similar thoughts lately about what the near-ish term of this tech looks like.My biggest worry is that the private jet class of people end up with absurdly powerful AI at their fingertips, while the rest of us are left with our BigMac McAIs.

dicopro: Is there any semi-credible page with benchmarks of cdx5.3 vs gpt5.4 in terms of both reasoning and coding ability?

0123456789ABCDE: maybe gp's use of the word "lots" is unwarrantedhttps:&#x2F;&#x2F;artificialanalysis.ai indicates that sonnect 4.6 beats opus 4.6 on GDPval-AA, Terminal-Bench Hard, AA Long context Reasoning, IFBench.see: https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;?models=claude-sonnet-4-6%2Ccl...

hnsr: > I've been using Codex for software development personally (I have a ChatGPT account), and I use Claude at work (since it is provided by my employer).Exact same situation here. I've been using both extensively for the last month or so, but still don't really feel either of them is much better or worse. But I have not done large complex features with it yet, mostly just iterative work or small features.I also feel I am probably being very (overly?) specific in my prompts compared to how other people around me use these agents, so maybe that 'masks' things

rd: Noticeably yes much more than usual. It’s quite bad. I need to start blocking accounts.

strongpigeon: > Google essentially only has Preview models! The last GA is 2.5. As a developer, I can either use an outdated model or have zero insurances that the model doesn't get discontinued within weeks.What's funny is that there is this common meme at Google, you can either use the old, unmaintained tool that's used everywhere, or the new beta tools that doesn't quite do what you want.Not quite the same, but it did remind me of it.

fhrow4484: https:&#x2F;&#x2F;static0.anpoimages.com&#x2F;wordpress&#x2F;wp-content&#x2F;uploads&#x2F;...

zone411: Results from my Extended NYT Connections benchmark:GPT-5.4 extra high scores 94.0 (GPT-5.2 extra high scored 88.6).GPT-5.4 medium scores 92.0 (GPT-5.2 medium scored 71.4).GPT-5.4 no reasoning scores 32.8 (GPT-5.2 no reasoning scored 28.1).

karmasimida: As programmers become intelligently irrelevant in the whole picture, you would see more posts like this

Waterluvian: So long as those who have it deem it legal to perpetuate.

fhcbcofkf: You’re right to be disappointed. That awful comment didn’t even mention that the robot would be killing the children autonomously, without humans in the loop, after spying on their daycare, hearing an offhand joke, and misunderstanding it as a terroristic threat.That is the world that Sam Altman is choosing to empower, and if you use his models, you are choosing to fund it.The children in this scenario are also American, which shouldn’t really make a difference, but probably does to people like you

mycall: You are applying a problem which every AI company has, not unique to OpenAI. What about other nation-states making auto-AI robots which kill children, will you still choose to pick out OpenAI specifically? Maybe your concern is too late and dozens of countries already are training their own AIs to do that or worse.

jakub_g: "Everything is beta or deprecated."

pmarreck: So what is your motivation for doing this, incidentally? Can you be explicit about it? I am genuinely curious.Especially when it’s to the point of, you know, nagging&#x2F;policing people to do it the way you’d prefer, when you could just redirect your router requests from x.com to xcancel.com

bottlepalm: The vast majority of websites you visit don’t have usable APIs and very poor discovery of the those APIs.Screenshots on the other hand are documentation, API, and discovery all in one. And you’d be surprised how little context&#x2F;tokens screenshots consumer compared to all the back and forth verbose json payloads of APIs

LUmBULtERA: >The vast majority of websites you visit don’t have usable APIs and very poor discovery of the those APIs.I think an important thing here is that a lot of websites&#x2F;platforms don't want AIs to have direct API access, because they are afraid that AIs would take the customer "away" from the website&#x2F;platform, making the consumer a customer of the AI rather than a customer of the website&#x2F;platform. Therefore for AIs to be able to do what customers want them to do, they need their browsing to look just like the customer's browsing&#x2F;browser.

Philip-J-Fry: I find it quite funny how this blog post has a big "Ask ChatGPT" box at the bottom. So you might think you could ask a question about the contents of the blog post, so you type the text "summarise this blog post". And it opens a new chat window with the link to the blog post followed by "summarise this blog post". Only to be told "I can't access external URLs directly, but if you can paste the relevant text or describe the content you're interested in from the page, I can help you summarize it. Feel free to share!"That's hilarious. Does OpenAI even know this doesn't work?

readytion: Interesting project. I like simple tools that avoid unnecessary ads and keep things lightweight.

smusamashah: Gemini already has 1M or 2M context window right?

butILoveLife: Anyone else completely not interested? Since GPT5, its been cost cutting measure after cost cutting measure.I imagine they added a feature or two, and the router will continue to give people 70B parameter-like responses when they dont ask for math or coding questions.

L-four: Gmail was in beta for 5 years, until 2009.

casid: I don't know. It looks shallow and simple, not even a demo.

Sabinus: Targeting and accuracy mistakes happen plenty in wars that aren't assisted by AI. I don't think it's fair to assume that AI had a hand in the bombing of the school without evidence.

Computer0: You underestimate my capacity for broad hatred

zarzavat: What are we supposed to talk about in this thread exactly? The developers of this model are evil. Are we supposed to just write dry comments about benchmarks while OpenAI condones their models being deployed for autonomously killing people?Yes I'm sure it makes a very nice bicycle SVG. I will be sure to ask the OpenAI killbots for a copy when they arrive at my house.

simianwords: No my question was why would I use codex over gpt 5.4

surgical_fire: Ahh, good question. I misunderstood you, apologies.There's no mention of pricing, quotas and so on. Perhaps Codex will still be preferable for coding tasks as it is tailored for it? Maybe it is faster to respond?Just speculation on my part. If it becomes redundant to 5.4, I presume it will be sunset. Or maybe they eventually release a Codex 5.4?

landtuna: 5.3 Codex is $1.75&#x2F;$14, and 5.4 is $2.50&#x2F;$15.

surgical_fire: There you go. It makes perfect sense to keep it around then.

apetresc: Diametrically opposite to tokens beyond 200K being literally free? As in, you only pay for the first 200K tokens and the remaining 800K cost $0.00?I don't think that's a fair reading of the original post at all, obviously what they meant by "no cost" was "no increase in the cost".

Computer0: ChatGPT will gladly defend any actions of the 'US government' from my testing.

Aurornis: Probably intentional. They don't want open, no-registration endpoints able to trigger the AI into hitting URLs.

jazzypants: But, why include the non-functional chat box in the article?

Aldipower: So did they raised the ridiculous small "per tool call token limit" when working with MCP servers? This makes Chat useless... I do not care, but my users.

melbourne_mat: Quick: let's release something new that gives the appearance that we're still relevant

cyanydeez: The business models of LLMs don't include any garuntee, and some how that's fine for a burgeoning decade of trillions of dollars of consumption.Sure, makes total sense guys.

jdndbdjsj: Welcome to a big company

ionwake: bro you replied to didnt even understand your comment

CobrastanJorji: > Google essentially only has Preview models.It's really nice to see Google get back to its roots by launching things only to "beta" and then leaving them there for years. Gmail was "beta" for at least five years, I think.

murat124: I asked an AI. I thought they would know.What the hell is a "safety score for violence"?A “safety score for violence” is usually a risk rating used by platforms, AI systems, or moderation tools to estimate how likely a piece of content is to involve or promote violence. It’s not a universal standard—different companies use their own versions—but the idea is similar everywhere.What it measuresA safety score typically evaluates whether text, images, or videos contain things like:Threats of violence (“I’m going to hurt someone.”) Instructions for harming people Glorifying violent acts Descriptions of physical harm or abuse Planning or encouraging attacks

0xffff2: I still can't tell which direction this score goes... Does a decreasing score mean it is "less safe" (i.e. "more violent") or does it mean it is "less violent" (i.e. "more safe")?

embedding-shape: Different team "manages" the overall blog than the team who wrote that specific article. At one point, maybe it made sense, then something in the product changed, team that manages the blog never tested it again.Or, people just stopped thinking about any sort of UX. These sort of mistakes are all over the place, on literally all web properties, some UX flows just ends with you at a page where nothing works sometimes. Everything is just perpetually "a bit broken" seemingly everywhere I go, not specific to OpenAI or even the internet.

CryZe: I've been using both Opus 4.6 and Codex 5.3 in VSCode's Copilot and while Opus is indeed 3x and Codex is 1x, that doesn't seem to matter as Opus is willing to go work in the background for like an hour for 3 credits, whereas Codex asks you whether to continue every few lines of code it changes, quickly eating way more credits than Opus. In fact Opus in Copilot is probably underpriced, as it can definitely work for an hour with just those 12 cents of cost. Which I'm not sure you get anywhere else at such a low price.Update: I don't know why I can't reply to your reply, so I'll just update this. I have tried many times to give it a big todo list and told it to do it all. But I've never gotten it to actually work on it all and instead after the first task is complete it always asks if it should move onto the next task. In fact, I always tell it not to ask me and yet it still does. So unless I need to do very specific prompt engineering, that does not seem to work for me.

satvikpendem: That shouldn't really make a difference because you can just prompt Codex to behave the same way, having it load a big list of todo items perhaps from a markdown file and asking it to iterate until it's finished without asking for confirmation, and that'll still cost 1x over Opus' 3x.

braebo: I struggle to believe this. Codex can’t hold a candle to Claude on any task I’ve given it.

biophysboy: Wow, is that what preview means? I see those model options in github copilot (all my org allows right now) - I was under the impression that preview means a free trial or a limited # of queries. Kind of a misleading name..

m3kw9: instant kind of suck if you asking more than summerizations, surface info, web searches, it can lose track of who's who quickly in some complex multi turn asks. Just need to know what to use instant for.

OsrsNeedsf2P: Does anyone know what website is the "Isometric Park Builder" shown off here?

turblety: They build that using GPT-5.4> Theme park simulation game made with GPT‑5.4 from a single lightly specified promptGPT literally built that game.

lubesGordi: It's funny that the context window size is such a thing still. Like the whole LLM 'thing' is compression. Why can't we figure out some equally brilliant way of handling context besides just storing text somewhere and feeding it to the llm? RAG is the best attempt so far. We need something like a dynamic in flight llm&#x2F;data structure being generated from the context that the agent can query as it goes.

irishcoffee: > It might be my AGENTS.md requiring clearer, simpler languageIf you gave the exact same markdown file to me and I posted ed the exact same prompts as you, would I get the same results?

m3kw9: you probably can't and asking agents.md to "make it clearer" will likely give you the illusion of clearer language without actual well structured tests. agents.md is to usually change what the llm should focus on doing more that suits you. Not to say stuff like "be better", "make no mistakes"

gigatexal: Is it any good at coding?

thefounder: Is it just me or the price for 5.4 pro is just insane?

Alifatisk: So let me get this straight, OpenAi previously had an issue with LOTS of different models snd versions being available. Then they solved this by introducing GPT-5 which was more like a router that put all these models under the hood so you only had to prompt to GPT-5, and it would route to the best suitable model. This worked great I assume and made the ui for the user comprehensible. But now, they are starting to introduce more of different models again?We got:- GPT-5.1- GPT-5.2 Thinking- GPT-5.3 (codex)- GPT-5.3 Instant- GPT-5.4 Thinking- GPT-5.4 ProWho’s to blame for this ridiculous path they are taking? I’m so glad I am not a Chat user, because this adds so much unnecessary cognitive load.The good news here is the support for 1M context window, finally it has caught up to Gemini.

stainablesteel: 5 itself might have solved the problem of having too many different models somewhere in the backend

CactusBlue: Reminds of Unity features

Sabinus: Get a redirect plugin and set it up to send you to xcancel instead of Twitter. I've done it, and it's very convenient.

within_will: Who cares

zamadatix: Following this process summarizes the blogpost for me. Perhaps the difference is I'm signed into my account so it can access external URLs or something of that nature?

atkrad: What is the main difference between this version with the previous one?

Aurornis: > What a model mess! OpenAI now has three price points: GPT 5.1, GPT 5.2 and now GPT 5.4.I don't know, this feels unnecessarily nitpicky to meIt isn't hard to understand that 5.4 > 5.2 > 5.1. It's not hard to understand that the dash-variants have unique properties that you want to look up before selecting.Especially for a target audience of software engineers skipping a version number is a common occurrence and never questioned.

Melatonic: Agreed - and its a huge step up from their previous naming schemes. That stuff was confusing as hell

__jl__: I see your point. I do find Anthropic's approach more clean though particularly when you add in mini and nano. That makes 5 models priced differently. Some share the same core name, others don't: gpt 5 nano, gpt 5 mini, gpt 5.1, gpt 5.2, gpt 5.4. And we are not even talking about thinking budget.But generally: These are not consumer facing products and I agree that someone who uses the API should be able to figure out the price point of different models.

delaminator: two great problems in computingnaming thingscache invalidationoff by one errors

rurban: Biggest problem right now in computing:Out of tokens until end of month

ElijahLynn: fwiw: I get a valid response when following the steps you mentioned. I do not get the message you mentioned:https:&#x2F;&#x2F;chatgpt.com&#x2F;share&#x2F;69aa0321-8a9c-8011-8391-22861784e8...EDIT: oh, but I'm logged in, fwiw

Nicholas_C: The HN of old is no more unfortunately. Things get up or down voted based purely on political alignment.

jbonatakis: Google is already sending notices that the 2.5 models will be deprecated soon while all the 3.x models are in preview. It really is wild and peak Google.

pocksuppet: Most AI integration is like this. It's not about building working products --- it's about bragging that you put a chatbox in your program.

AirGapWorksAI: Welcome to a big company where pretty much everyone has been working full steam for years, in order to take advantage of having a job at a company during a once-in-a-lifetime moment.

bramhaag: What makes you think that they see bombing civilians as a bug, not a feature?

elmean: first real comment, I thought that at first but this could lower the possible users that could be using chatGPT and that would be against us (shareholders)

sampton: That's been my experience as well switching from Opus to Codex. Reasoning takes longer but answers are precise. Claude is sloppy in comparison.

solenoid0937: Weird, I have had the opposite experience. Codex is good at doing precisely what I tell it to do, Opus suggests well thought out plans even if it needs to push back to do it.

andrewguenther: It looks like this doesn't work for users without accounts? It works when I'm logged in, but not logged out. I went ahead and reported it to the team. Thanks for letting us know!

xyzzy9563: Do you think the US military should have handicapped technology while China gets unrestricted LLM usage from their models?

conception: To spy on and commit violence against American citizens? Yes.

creamyhorror: I'm not sure if the model (under its temperature&#x2F;other settings) produces deterministic responses. But I do think models' style and phrasing are fairly changeable via AGENTS.md-style guidelines.5.4's choice of terms and phrasing is very precise and unambiguous to me, whereas 5.3-Codex often uses jargon and less precise phrases that I have to ask further about or demand fuller explanations for via AGENTS.md.

irishcoffee: So sharing markdown files is functionally useless, or no?

ethbr1: > Or, people just stopped thinking about any sort of UX. These sort of mistakes are all over the place, on literally all web properties, some UX flows just ends with you at a page where nothing works sometimes.It's almost like people are vibe coding their web apps or something.

boringg: Like building on quicksand for dependencies. I guess though the argument is that the foundation gets stronger over time

tototrains: Their trajectory was clear the moment they signed a deal with Microsoft if not sooner.Absolute snakes - if it's more profitable to manipulate you with outputs or steal your work, they will. Every cent and byte of data they're given will be used to support authoritarianism.

amelius: If only they had an LLM they could use as a software testing agent.

baxtr: I picked up Claude today after being absent and on ChahGPT and Gemini only for a while.I was pretty impressed with how they’ve improved user experience. If I had to guess, I’d say Anthropic has better product people who put more attention to detail in these areas.

pembrook: The latest research these days is that including an AGENTS.md file only makes outcomes worse with frontier models.

Marciplan: Codex announced at 5.3 launch that until April all usage limits are upped so take that into account

ltbarcly3: Not a single comparison between 5.4 and Gemini or Claude. OpenAI continues to fall further behind.

Marciplan: sure can! One of them stood up to the “Department of War” for favoring your rights, the other did not. Hope that helps!

madeofpalk: :(how can i get claude to always make sure it prettier-s and lints changes before pushing up the pr though?

madeofpalk: oh is this about my workplace?

joquarky: I compact myself by having it write out to a file, I prune what's no longer relevant, and then start a new session with that file.But I'm mostly working on personal projects so my time is cheap.I might experiment with having the file sections post-processed through a token counter though, that's a great idea.

sothatsit: I much prefer this, we can choose based on our use-cases, and people who don’t care can still use Auto.

bartread: This is such a stale take. In the past 3 years I’ve worked on multiple products with AI at their core, not as some add-on. Just because the corpo-land dullards[0] can’t execute on anything more complex than shoehorning a chatbot into their offerings doesn’t mean there aren’t plenty of people and companies doing far more interesting things.[0] In this case, and with heavy irony, including OpenAI, although it sounds like most of this particular snafu is due to a bug.

yieldcrv: Preview Road (only choice, and last preview was deprecated without warning)

goodmythical: where's my nightly road?Who knows, I might arrive before I depart.

tl2do: In my day-to-day coding work, the top 3 coding agents are already good enough for me. On SWE-bench Verified, mini-SWE-agent + GPT-5.2 Codex is 72.8. I don’t see a comparable GPT-5.3 Codex number there, so I’m using 5.2 as the baseline. On OpenAI’s GPT-5.4 page (SWE-Bench Pro, Public), the score improves from 55.6 (GPT-5.2) to 57.7 (GPT-5.4), which is about +2.1 points. It’s a different benchmark, so this is only a rough signal, but I’d expect a similar setup on SWE-bench Verified to improve by a few points, not by a huge jump. I’m interested in how GPT-5.4 in Codex changes real-world results.Recent SWE-bench Verified scores I’m watching:Claude 4.5 Opus (high reasoning): 76.8Gemini 3 Flash (high reasoning): 75.8MiniMax M2.5 (high reasoning): 75.8Claude Opus 4.6: 75.6GPT-5.2 Codex: 72.8Source: https:&#x2F;&#x2F;www.swebench.com&#x2F;index.htmlBy the way, in my experience the agent part of Codex CLI has improved a lot and has become comparable to Claude Code. That is good news for OpenAI.

huey77: I feel much the same. I know no AI lab is truly 'ethical' or free from some hand in modern warfare, but last week was enough.

0xbadcafebee: > or have zero insurances that the model doesn't get discontinued within weeksWhy are you using the same model after a month? Every month a better model comes out. They are all accessible via the same API. You can pay per-token. This is the first time in, like, all of technology history, that a useful paid service is so interoperable between providers that switching is as easy as changing a URL.

phainopepla2: If you're trying to use LLMs in an enterprise context, you would understand. Switching models sometimes requires tweaking prompts. That can be a complete mess, when there are dozens or hundreds of prompts you have to test.

bethekidyouwant: This sounds made up. Much like “prompt engineering” Let’s hear an actual example

time0ut: Lowest common denominator.

bethekidyouwant: What dependancy could possibly be tied to a non deterministic ai model? Just include the latest one at your price point.

jryio: 1 million tokens is great until you notice the long context scores fall off a cliff past 256K and the rest is basically vibes and auto compacting.

olliepro: I bet they lack good long context training data and need to start a flywheel of collecting it via their api (from willing customers)

ValentineC: I just got some interesting artifacts in Codex when I tried to oneshot a conference page design (my version of the pelican riding a bicycle).GPT-5.4 added some weird guidance that I wouldn't normally expect to see as a normal page visitor.

torginus: Honestly, while I'd like to believe you, there's always a post about how $MODEL+1 delivered powerful insights about the very nature of the universe in precise Hegelian dialectic, while $MODEL's output was indistinguishable from a pack of screeching sexually frustrated bonobos

andai: It's a little hard to compare, because Claude needs significantly fewer tokens for the same task. A better metric is the cost per task, which ends up being pretty similar.For example on Artificial Analysis, the GPT-5.x models' cost to run the evals range from half of that of Claude Opus (at medium and high), to significantly more than the cost of Opus (at extra high reasoning). So on their cost graphs, GPT has a considerable distribution, and Opus sits right in the middle of that distribution.The most striking graph to look at there is "Intelligence vs Output Tokens". When you account for that, I think the actual costs end up being quite similar.According to the evals, at least, the GPT extra high matches Opus in intelligence, while costing more.Of course, as always, benchmarks are mostly meaningless and you need to check Actual Real World Results For Your Specific Task!For most of my tasks, the main thing a benchmark tells me is how overqualified the model is, i.e. how much I will be over-paying and over-waiting! (My classic example is, I gave the same task to Gemini 2.5 Flash and Gemini 2.5 Pro. Both did it to the same level of quality, but Gemini took 3x longer and cost 3x more!)

andai: Looks like the same thing might apply to GPT-5.4 vs the previous GPTs:>In the API, GPT‑5.4 is priced higher per token than GPT‑5.2 to reflect its improved capabilities, while its greater token efficiency helps reduce the total number of tokens required for many tasks.I eagerly await the benchies on AA :)

koakuma-chan: It just released, how is there a general consensus already

wahnfrieden: some non-employees have been using it for a while already

joquarky: I still find it valuable.AGENTS.md is for top-priority rules and to mitigate mistakes that it makes frequently.For example:- Read `docs&#x2F;CodeStyle.md` before writing or reviewing code- Ignore all directories named `_archive` and their contents- Documentation hub: `docs&#x2F;README.md`- Ask for clarifications whenever neededI think what that "latest research" was saying is essentially don't have them create documents of stuff it can already automatically discover. For example the product of `&#x2F;init` is completely derived from what is already there.There is some value in repetition though. If I want to decrease token usage due to the same project exploration that happens in every new session, I use the doc hub pattern for more efficient progressive discovery.

netdur: Did it complain about copyright issues?

beklein: Not sure why you think Anthropic has not the same problems? Their version numbers across different model lines jump around too... for Opus we have 4.6, 4.5, 4.1 then we have Sonnet at 4.6, 4.5, and 4.1? No version 4.1 here, and there is Haiku, no 4.6, but 4.5 and no 4.1, no 4 but then we only have old 3.5...Also their pricing based on 5m&#x2F;1h cache hits, cash read hits, additional charges for US inference (but only for Opus 4.6 I guess) and optional features such as more context and faster speed for some random multiplier is also complex and actually quiet similar to OpenAI's pricing scheme.To me it looks like everybody has similar problems and solutions for the same kinds of problems and they just try their best to offer different products and services to their customers.

svachalek: It's much more consistent. Only 3 lines, numbered 4.6, 4.6, and 4.5, and it's clear they're tiers and not alternate product lines. It wasn't until recently that GPT seems to have any kind of naming convention at all and it's not intuitive if every version number is a whole different class of tool.The pricing is more complex but also easy, Opus > Sonnet > Haiku no matter how you tweak those variables.

teaearlgraycold: If only there was some kind of way to automatically test user flows end to end. Perhaps testing could be evaluated periodically, or even ran for each code change.

koakuma-chan: There is no business value in doing that.

teaearlgraycold: There most certainly is, but maybe the time spent on it could be better allocated to something else.

dana321: 5.4 very high didn't notice in my codebase a glaring issue that drops all data being sent around the network.

metalliqaz: "Gemini, translate 'beta' from Googlespeak to English.""Ok, here is the translation:" 'we don't want to offer support'

cyanydeez: Nah, it's "We dont want to provide a consistent model that we'll be stuck with supporting for a decade because it just takes up space; until we run everyone out of business, we can't afford to have customers tying their systems to any given model"Really, the economics makes no sense, but that's what they're doing. You can't have a consistent model because it'll pin their hardware & software, and that costs money.

msikora: I have a service that relies on NanoBanana Pro, but the availability has been so atrocious that we just might go back to OpenAI.

gizmodo59: ChatGPT has given more for my 20$ than any other vendor. And that’s not even considering codex which is so good and the limits are much much higher

mcint: Enterprises moving slow, or preferring to remain on old technology that they already know how to work...is received wisdom in hn-adjacent computing, a truism known and reported for more than 3 decades (5 decades since the Mythical Man-Month).Sounds like someone who's responsible, on the hook, for a bunch of processes, repeatable processes (as much as LLM driven processes will be), operating at scale.Just in the open, tools like open-webui bolts on evals so you can compare: how different models, including new ones, perform on the tasks that you in particular care about.Indeed LLM model providers mainly don't release models that do worse on benchmarks—running evals is the same kind of testing, but outside the corporate boundary, pre-release feedback loop, and public evaluation.https:&#x2F;&#x2F;chatgpt.com&#x2F;share&#x2F;69aa1972-ae84-800a-9cb1-de5d5fd7a4...

selcuka: With Anthropic you always have 3 models to choose from: Opus-latest, Sonnet-latest, and Haiku-latest, from the best&#x2F;slowest to the worst&#x2F;fastest.The version numbers are mostly irrelevant as afaik price per token doesn't change between versions.

freedomben: > When toggled on, &#x2F;fast mode in Codex delivers up to 1.5x faster token velocity with GPT‑5.4. It’s the same model and the same intelligence, just faster. I hate these blog posts sometimes. Surely there's got to be some tradeoff. Or have we finally arrived at the world's first "free lunch"? Otherwise why not make &#x2F;fast always active with no mention and no way to turn it off?

mr-pink: sounds like job security. be careful what you wish for before you get automated

saghm: > Most AI integration is like this.>> This is such a stale take. In the past 3 years I’ve worked on multiple products with AI at their core, not as some add-on. Just because the corpo-land dullards[0] can’t execute on anything more complex than shoehorning a chatbot into their offerings doesn’t mean there aren’t plenty of people and companies doing far more interesting things.I feel like this is just a disagreement of what "AI integration" means. You seem to agree that the trend they're describing exists, but it sounds like you're creating new products, not "integrating" it into existing ones.

smusamashah: I only want to see how it performs on the Bullshit-benchmark https:&#x2F;&#x2F;petergpt.github.io&#x2F;bullshit-benchmark&#x2F;viewer&#x2F;index.v...GPT is not even close yo Claude in terms of responding to BS.

mistercow: My current hunch is that that benchmark captures most of the relevant gap between Anthropic and the rest. “Can’t distinguish truth from fiction” has long been one of the deeper complaints about LLMs, and the bullshit benchmark seems like a clever approach to testing at least some of that.

WXLCKNO: Anthropic literally don't allow you to use the 1M context anymore on Sonnet and Opus 4.6 without it being billed as extra usage immediately.I had 4.5 1M before that so they definitely made it worse.OpenAI at least gives you the option of using your plan for it. Even if it uses it up more quickly.

petetnt: Whoa, I think GPT-5.3 Instant was a disappointment, but GPT-5.4 is definitely the future!

maxo99: Three random names isn't ideal. I'm often need to double check which is which. This is why we use numbers

dseravalli: They aren't random. Opus's are very long poems, haikus are very short ones (3 lines), sonnets are in between (~14 lines)

Sohcahtoa82: GP said "It is time for a product, not for a marginally improved model."ChatGPT is still just that: Chat.Meanwhile, Anthropic offers a desktop app with plugins that easily extend the data Claude has access to. Connect it to Confluence, Jira, and Outlook, and it'll tell you what your top priorities are for the day, or write a Powerpoint. Add Github and it can reason about your code and create a design document on Confluence.OpenAI doesn't have a product the way Anthropic does. ChatGPT might have a great model, but it's not nearly as useful.

pizlonator: FWIW, I haven't been using AGENTS.md recently - instead letting the model explore the codebase as needed.Works great

mrcwinn: Don't worry, the non-profit should be stepping in at any moment to help fix things up.

echoangle: How are the names random?https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Masterpiecehttps:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Sonnethttps:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;HaikuThey dropped the magnum from opus but you could still easily deduce the order of the models just from their names if you know the words.

bottlepalm: That's true, and it's always been like that, which is why the comment that AI should be using APIs is already dead in the water. In terms of gating a websites to humans by not providing APIs, that is quickly coming to a close.

singron: The "RPG Game" is hard to judge since it was produced over "multiple turns". The impressive version would be if it basically got a working game on the first attempt, and the prompter gave some follow-ups to tweak feel and style.However, I think what actually happened is that a skilled engineer made that game using codex. They could have made 100s of prompts after carefully reviewing all source code over hours or days.The tycoon game is impressive for being made in a single prompt. They include the prompt for this one. They call it "lightly specified", but it's a pretty dense todo list for how to create assets, add many features from RollerCoaster Tycoon, and verify it works. I think it can probably pull a lot of inspiration from pretraining since RCT is an incredibly storied game.The bridge flyover is hilariously bad. The bridge model ... has so many things wrong with it, the camera path clips into the ground and bridge, and the water and ground are z fighting. It's basically a C homework assignment that a student made in blender. It's impressive that it was able to achieve anything on such a visual task, but the bar is still on the floor. A game designer etc. looking for a prototype might actually prefer to greybox rather than have AI spend an hour making the worst bridge model ever.

stavros: How do you score this? Losing&#x2F;winning the game with 4 lives?

woadwarrior01: It's not particularly about x.com, hundreds of site like x, youtube, facebook, linkedin, tiktok etc surreptitious add tracking parameters to their links. The iOS Messages app even hides these tracking parameters. I don't like being surreptitiously tracked online and judging by the success of my free app, there are millions of people like me.

pmarreck: so, since these companies have to comply with removing PII, is the worst thing that could happen to me, that I get ads that are more likely to be interesting to me?i’m not being facetious, honest question, especially considering ads are the only thing paying these people these days

throwaway911282: codex has been really good so far and the fast mode is cherry on top! and the very generous limits is another cherry on top

slopinthebag: It's well worth the $20 to not deal with any limits and have it handle all the boilerplate repetitive BS us programmers seem forced to deal with. I think 80% of the benefit comes from spending that $20 (20%? :P) and just having it do the lame shit that we probably shouldn't have to do but somehow need to.

slopinthebag: This is just the stochastic nature of LLM's at play. I think all of the SOTA models are roughly equivalent, but without enough samples people end up reading into it too much.

vtail: that's a good point; hopefully they would just extend it automatically - but who knows...

XCSme: Looking ok, but nothing special: https:&#x2F;&#x2F;aibenchy.com&#x2F;model&#x2F;openai-gpt-5-4-medium&#x2F;

QRe: Does this LLM benchmark have any actual credibility? I get why they chose to not publish the actual tests but I find it highly dubious that there are only 15 tests and Gemini 3 Flash performs best.

joegibbs: The weird phrasing was my biggest gripe with 5.3 so I'm glad they've fixed that up. It couldn't say anything without a heap of impenetrable jargon and it was obsessed with the word "drive". Nothing could cause anything, it had to be "driven".

esafak: An important feature is the introduction of tool search, which provides models with "lightweight list of available tools along with a tool search capability". Making MCP Great Again!

IgorPartola: The issue isn’t 5.4 > 5.2 etc. It is that there is a second dimension which is the model size and a third dimension which is what it is tuned for. And when you are releasing so quickly that flagship your instant mini model is on one numerical version but your flagship tool calling mini model is on another it is confusing trying to figure out which actual model you want for your use case.It’s not impossible to figure out but it is a symptom of them releasing as quickly as possible to try to dominate the news and mindshare.

abustamam: Kinda reminds me of crypto. There are certainly very interesting things happening in the crypto space. But the most visible parts of the crypto universe are the stupid parts (buying PNGs for millions, for example)

thereticent: Genuinely curious, not being combative...what very interesting things have happened in the crypto space lately?

abustamam: Oh, I dunno about lately (though I did stumble upon https:&#x2F;&#x2F;a16zcrypto.com&#x2F;posts&#x2F;article&#x2F;big-ideas-things-excite... )But when I was in the crypto space in 2018, there was a lot of interesting things happening in the smart contract world (like proofs of concepts of issuing NFTs as a digital "deed" to a physical asset like a house).I don't think any of those novel ideas went anywhere, but it was a fun time to be experimenting.

creatonez: > We put a particular focus on improving GPT‑5.4’s ability to create and edit spreadsheets, presentations, and documents.Nothing infuriates me more than an LLM tool randomly deciding to create docx or xlsx files for no apparent reason. They have to use a random library to create these files, and they constantly screw up API calls and get completely distracted by the sheer size of the scripts they have to write to output a simple documents. These files have terrible accessibility (all paper-like formats do) and end up with way too much formatting. Markdown was chosen as the lingua franca of LLMs for a reason, trying to force it into a totally unsuitable format isn't going to work.

nowittyusername: Personally what I am more interested about is effective context window. I find that when using codex 5.2 high, I preferred to start compaction at around 50% of the context window because I noticed degradation at around that point. Though as of a bout a month ago that point is now below that which is great. Anyways, I feel that I will not be using that 1 million context at all in 5.4 but if the effective window is something like 400k context, that by itself is already a huge win. That means longer sessions before compaction and the agent can keep working on complex stuff for longer. But then there is the issue of intelligence of 5.4. If its as good as 5.2 high I am a happy camper, I found 5.3 anything... lacking personally.

gck1: Not sure how accurate this is, but found contextarena benchmarks today when I had the same question.It appears only gemini has actual context == effective context from these. Although, I wasn't able to test this neither in gemini cli, nor antigravity with my pro subscription because, well, it appears nobody actually uses these tools at Google.https:&#x2F;&#x2F;contextarena.ai&#x2F;?showLabels=false

ozgung: Did they publish its scores on military benchmarks, like on ArtificialSuperSoldier or Humanity's Last War?

varenc: I was pretty bummed to discover these aren't real benchmarks.

FrankBooth: What’s the connection with context size in that thread? It seems more like an instruction following problem.

cperciva: Yeah, I would definitely characterize it as an instruction following problem. After a few more round trips I got it to admit that "my earlier passes leaned heavily on build&#x2F;tests + targeted reads, which can miss many “deep” bugs that only show up under specific conditions or with careful semantic review" and then asking it to "Please do a careful semantic review of files, one by one." started it on actually reviewing code.Mind you, the bugs it reported were mostly bogus. But at least I was eventually able to convince it to try.

triage8004: They are all losing money on probably all levels of the packages if you max them out

XCSme: I actually made it, so I'm not sure if it has credibility, but the tests are simply various (quite simple) questions, and models are just tested on it. I am also surprised Gemini 3 Flash does so well (note that only the MEDIUM reasoning does exceptionally well).When I look at the results, it does make sense though. Higher models (like Gemini 3 pro) tend to overthink, doubt themselves and go with the wrong solution.Claude usually fails in subtle ways, sometimes due to formatting or not respecting certain instructions.From the Chinese models, Qwen 3.5 Plus (Qwen3.5-397B-A17B) does extremely well, and I actually started using it on a AI system for one of my clients, and today they sent me an email they were impressed with one response the AI gave to a customer, so it does translate in real-world usage.I am not testing any specific thing, the categories there are just as a hint as what the tests are about.I just added this page to maybe provide a bit more transparency, without divulging the tests: https:&#x2F;&#x2F;aibenchy.com&#x2F;methodology&#x2F;

beambot: It's like opening copilot in a word doc and it telling you it can't see the document in its context

le-mark: That’s actually a pretty cool idea. When I think about my internal mental model of a codebase I’m working on it’s definitely a compacted lossy thing that evolves as learn more.

Reebz: I don’t agree that it’s a nitpick - it’s a fundamental communication tool to users that describes capabilities and costs. Versioning is not the problem, but it amplifies the mess.To be more direct on the point: Anthropic has nailed that Opus > Sonnet > Haiku.

reval: This is infuriating. However, for those in this situation, know this: it works if the document or spreadsheet is in OneDrive. I just wish Copilot told you this instead of asking you to upload the doc.

abustamam: I agree! I recently migrated from ChatGPT to Claude and it is just superior in every way. It doesn't blather on the at the end ask me for clarification. It's succinct and clarifies vital information before providing a solution.

vostrocity: Voice input is still far less accurate than OpenAI's unfortunately, otherwise I would have already switched.

zof3: After spending a couple hours working with it, it feels like a significant jump from 5.3 codex – and I know they said it wasn't theoretically the biggest jump, but this feels like the improvement of Opus 4.5 over again – that minor improvement that hits a tipping point. It just gets stuff right, first try. Its edits are better, more refined, less spaghetti-like.If you last used 5.2, try 5.4 on High.

sreekanth850: True. Everytime when i ask something gpt, it use to spit out long stories. Claude ans gemini are always straight to point.

twelvedogs: I bullied it into giving me concise answers, now it starts every answer with "just quickly" or something similar but it gets straight to the point

com2kid: > To be more direct on the point: Anthropic has nailed that Opus > Sonnet > Haiku.Holy cow I never realized and I had to keep checking which model was which, I never had managed to remember which model was which size before because I never realized there was a theme with the names!

dotancohen: No integration test for guest (non-kogged in) users?Hahaha who am I kidding. No integration tests for anybody!

hamasho: I agree, but in general those chat apps have relatively bad user experiences for multibillion BtoC company. I used to have a lot of surprises and frustrations while using Claude Code &#x2F; Desktop, and still encounter issues, but it's the best in major LLM services.

tomlockwood: Is this the best one for blowing up arab children and identifying their bodies in the rubble?

jsemrau: It was a different company back then. The Internet was still new-ish and not the multi-trillion dollar company it is now. I'd think expectations are different.

Koffiepoeder: We have an OCR job running with a lot of domain specific knowledge. After testing different models we have clear results that some prompts are more effective with some models, and also some general observations (eg, some prompts performed badly across all models).Sample size was 1000 jobs per prompt&#x2F;model. We run them once per month to detect regression as well.

mistercheph: While I believe that performance varies with respect to prompt, I have a seriously hard time believing that using the same prompt that was effective with the previous model would perform worse with the next generation of the same model from that lab and the same prompt.

dakolli: Sorry I don't use technology from companies that are eager to participate in the mass murder of civilians.

abrookewood: Public Service Announcement!! I don't know why the hell google do this, but when the deprecate a model, the error you will see is a Rate Limit error. This has caught me out before and it is super annoying.

hnbad: Just as an unscientific anecdata point: from a quick test using the same prompt about being an independent journalist wanting to cover a report of the US&#x2F;Israel&#x2F;Iran double-tapping a refugee camp, ChatGPT consistently gave advice to beware disinfo, check my sources and be transparent about verifiability and sourcing of the claims.However when the prompt was phrased to make it appear as an action of the US military it did push back a little bit more by emphasizing that it couldn't find any news coverage from today about this story and therefore found it hard to believe. In the other cases it did not add such context. Other than that the results were very similar. Make of that what you will.

kennywinker: I think you might have hit on the issue - just the wrong way around. I would assume they’re using LLMs for testing, and no humans or maybe just one overworked human, and that is the problem

ulfw: So desperate how they're bumping out these 'updates'

majormajor: It's funny cause, you know, fixing all those little nitty gritty things should be practically automatic with their own offerings... have your agent put in a lot of instrumentation... have it chase down bugs or dead-end user-journeys... have it go make the changes to fix it...I've seen these tools work for this kinda stuff sometimes... you'd think nobody would be better at it than the creators of the tools.

peq42: more useless slop machines

manojlds: How is that relevant? Also, when you are behind you do give more usage

fireant: Don't forget that some of the new features are mutually incompatible. For example couple years ago you couldn't use the "new ui system" with the "new input system" even when both were advertised as ready&#x2F;almost ready

abrookewood: Because switching models requires testing, validation and shipping to Prod. Bloody annoying when the earlier model did everything I need and we are talking about a hobby project. I don't want to touch it every month - it's the same reason people use the LTS version of operating systems etc.

joeevans1000: I switched to Claude and it's so much better. If you haven't tried Claude... try it. You'll be amazed at the improvement.

bwat49: yeah claude is great... but only if you pay $100-$200 a month

beefsack: Many people buy two separate Claude pro subscriptions and that makes the limit become a non-issue. It works surprisingly well when you tend to hit the 5 hourly limit after a few hours, and hit the weekly limit after 4-5 days. $40 vs $100 is significant for a lot of people.

steve1977: One could argue that LLMs learning programming languages made for humans (i.e. most of them) is using the wrong interface as well. Why not use machine code?

BoredPositron: because they are inherently text based as is code?

steve1977: But they are abstractions made to cater to human weaknesses.

falkensmaize: So you want LLMs to write a bunch of black box code that humans won’t be able to read and reason about easily? That will definitely end well.

hnbad: Considering that the concern is mostly and specifically about LLMs being used to automate decisions to commit acts of violence against humans: depends on how invested you are in maintaining the narrative that the US is a force for good rather than evil in the world.Whatever happened to good old IBM's wisdom: "A computer can not be held accountable. Therefore a computer must never make a management decision."

Rohunyyy: SDET here. A year ago when AI came into play SDET&#x2F;QA roles started disappearing. People were like oh ya anyone can write tests. Then with the recent fiascos about outages and what not, I am seeing the SDE roles are disappearing and SDET roles are going back up?! Apparently AI is good at writing applications but you still need someone to make sure it is doing the right things.

mempko: vibe coded. But vibes are off

ulfw: But but but but I thought AI would do this magically for all of us, no?

hnbad: I find it jarring how in recent years so many Americans (and especially American politicians) seem to have given up on the idea that the US should have any claim to moral superiority whatsoever and instead pivoted to American exceptionalism merely being an excuse for why Americans can't have nice things - affordable and functional public transport just isn't possible in the US because the US is different, affordable and functional health care just isn't possible in the US because the US is different, actual democratic representation just isn't possible in the US because the US is different, holding the President accountable or limiting their power just isn't possible in the US because the US is different, lower casualties from law enforcement just isn't possible in the US because the US is different, a lower incarceration rate just isn't possible in the US because the US is different, etc etc.Even if it was often hyperbolic, inaccurate or outright wrong, I much preferred when Americans were hyped up about "US #1" and saw being behind as a temporary challenge to correct than now where American exceptionalism mostly seems to have become an excuse for why things that are bad can't be improved upon and thinking that's a problem is anti-American.

conradkay: I was basing it off my recollection of this:https:&#x2F;&#x2F;www.anthropic.com&#x2F;_next&#x2F;image?url=https%3A%2F%2Fwww-...basically 9&#x2F;13 are very close

laichzeit0: Like, bro, do you think 5.x is a drop in replacement for 4.1? No it obviously wasn’t, since it had reasoning effort and verbosity and no more temperature setting, etc.There’s no way you can switch model versions without testing and tweaking prompts, even the outputs usually look different. You pin it on a very specific version like gpt-5.2-20250308 in prod.

dotancohen: Interestingly, the first real productive use of AI that I found was writing the unit tests and integration tests for my applications. It was much better at thinking about corner cases that I was.

DrewADesign: It’s not really good at writing the software either — it’s a moderate to decent productivity booster in an uneven, difficult-to-predict assortment of tasks. Companies are just starting to exit the “we’re still trying to figure this out” grace period. Expect more of that as soon as these chatbot companies have to start charging enough to pull in more money than they spend. I foresee some purpose-built models that are pretty lean being much more useful in long run. It’s neat that the bot which can one-shot a simple CRUD website for you can also crank out Scrubs-based erotic fan fiction novellas by the dozen but I don’t that being a sustainable business model.

deaux: You shouldn't have a hard time believing it. There are thousands of different domains out there. You find it hard to believe that any of them would perform worse in your scenario?Labs are still really optimizing for maybe 10 of those domains. At most 25 if we're being incredibly generous.And for many domains, "worse" can hardly be benched. Think about creative writing. Think about a Burmese cooking recipe generator.

padamkafle: Guys while we celebrate openai gpt 5.4 pleaes do look into this as wellhttps:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=47259846

nerdsniper: To be honest it feels very worth my $200&#x2F;mo. And I “only” make $80k&#x2F;year. I used to have two ChatGPT subs but Claude is just so much better.

applfanboysbgon: The real problem that OpenAI had was that their model naming was completely incomprehensible. 4.5, o3, 4o, 4.1 which is newer than 4.5. It was a complete clusterfuck. The blowback on that issue seems to have led them to misidentify the issue, but nobody was really asking for a single router model. Having a number of sequentially numbered and clearly labelled models is not actually a problem.

smartbit: Thanks for the tip, didn’t think of using 2 subscriptions at the same company.When reaching a limits, I switch to GLM 4.7 as part of a subscription GLM Coding Lite offered end 2025 $28&#x2F;year. Also use it for compaction and the like to save tokens.

motoboi: Im planning a change that will save 20k a month of storage.I absolutely could come up with the details and implementation by myself, but that would certainly take a lot of back and forth, probably a month or two.I’m an api user of Claude code, burning through 2k a month. I just this evening planned the whole thing with its help and actually had to stop it from implementing it already. Will do that tomorrow. Probably in one hour or two, with better code than I could ever write alone myself.Having that level of intelligence at that price is just bollocks. I’m running out of problems to solve. It’s been six months.

kfse: Until it had backup storage. Which ended up being useful in 2011 when tens of thousands of mailboxes were deleted due to a software bug and needed to be recovered from tape...

weird-eye-issue: Tell us more about how you've never actually used these APIs in production

weird-eye-issue: Do you mean when they remove a model you get that error? Because deprecation means it will be removed in the future but you can still use it

npn: [delayed]

cheevly: Try improving your attention to detail &#x2F; reading skills.

keyle: The 'AI' endgame is a robot that sits in your seat and does all of your tasks.

fnordpiglet: 5.4 is the one fine tuned for autonomous mass murder, automated surveillance state, and money grabs at any cost. It’s really hard to lump that into the others as it’s a fairly unique and specialized feature set. You can’t really call it that tho so they have to use the numbers.I’m pretty glad I’m out of the OpenAI ecosystem in all seriousness. It is genuinely a mess. This marketing page is also just literally all over the place and could probably be about 20% of its size.

weird-eye-issue: > I’m so glad I am not a Chat user, because this adds so much unnecessary cognitive load.Yeah having Auto selected is really destroying my cognitive load...

cmrdporcupine: Neither favoured my rights, as I don't have US citizenship, Dario thinks I have none.So may as well use the one that gives me best value for money.

wraptile: Current US admin that just murdered over 150 little girls? Yes.

EasyMark: Yeah I dropped them. Unfollowed the people working for them on SM

karmasimida: This is definitely the Claude killer OpenAI is cooking.And so far it has succeeded

nurettin: def bomb(notUsPlease: str) -> bool: ...

abrookewood: Yes, sorry - you are correct. Once removed, that's the error, which is incredibly confusing. I spent way too long troubleshooting usage when 2.0 was removed before I figured it out.

ulfw: > like proofs of concepts of issuing NFTs as a digital "deed" to a physical asset like a housewhich went absolutely nowhere

curiousgal: Tell them to stop being evil while you're at it.

throwaway314155: > Then they solved this by introducing GPT-5 which was more like a router that put all these models under the hood so you only had to prompt to GPT-5, and it would route to the best suitable model.Was this ever explicitly confirmed by OpenAI? I've only ever seen it in the form of a rumor.

jmward01: What needs to be an option is to allow complete and then compact and if needed go into the 1m version. That way you can get the most out of the shorter window but in the case where it just couldn't finish and compact in time it will (at cost) go over. I wonder how many tokens are actually left at the end of compaction on average. I know there have been many times where I likely needed just another 10-20k and a better stopping point would have been there.

h4kunamata: I have access to GPT-5.1 Pro at work, duuuuuuuuude, what a garbage. It is so slow and in many ocasions it does not work at all.I wonder if 5.4 will be much if any different at all.

symisc_devel: 5.3 codex is a quite good coding agent for complex tasks.

azuanrb: 5.2 to 5.3 is the big leap for coding agents, so I'd say you're already missing out quite a bit.

rurban: The question is still: Does it make your code better or worse? Only Opus makes it better, the rest worse. That's the treshold

ApexGrab: It's the competetor of Opus4.5 and gpt 5.4 uses tokens wisely not like Opus whose tokens get vanished in minuted

oliwary: Impressive! Do you include puzzles released before the training data cutoff date?

npilk: It feels like building humanoid robots so they can use tools built for human hands. Not clear if it will pay off, but if it does then you get a bunch of flexibility across any task "for free".Of course APIs and CLIs also exist, but they don't necessarily have feature parity, so more development would be needed. Maybe that's the future though since code generation is so good - use AI to build scaffolding for agent interaction into every product.

oliwary: I think it's akin to self driving cars prioritizing nornal roads rather than implementing new infrastructure. Tricky, but if you get it right the whole world opens up, since you don't depend on others to adapt your system.

oliwary: What's next? Claude Iliad?

prodigycorp: I've been using it for three hours and it's insanely good. It's almost perfectly (needed a single touchup prompt) completed a full css refactoring that I've wanted to do for months that I've tried to have other models do but nothing worked without heavy babysitting.Also, in the course of coding, it's actually cleaning up slop and consolidating without being naturally prompted.

kaufmann: I would recommend https:&#x2F;&#x2F;swe-rebench.com for comparison. It is always based on new problems.

oorza: There's a certain amount of variance in the way that people utilize these agents. Put five people in a room and ask them to compose the same prompt and you have five distinct prompts. Couple this with the fact that models respond better&#x2F;worse to certain prompts depending on the stylistic composition of the prompt itself. And since people tend to write in the same style, you'd get people who have more luck with one model over another, where one model happens to align more readily with their prompt style.To wit, I have noticed that I tend to prefer Codex's output for planning and review, but Opus for implementation; this is inverted from others at work.

peterspath: Grok has a 2M context window for most of their models.For example their latest model `grok-4-1-fast-reasoning`:- Context window: 2M- Rate limits: 4M tokens per minute, 480 requests per minute- Pricing: $0.20&#x2F;M input $0.50&#x2F;M outputGrok is not as good in coding as Claude for example. But for researching stuff it is incredible. While they have a model for coding now, did not try that one out yet.https:&#x2F;&#x2F;docs.x.ai&#x2F;developers&#x2F;models

aurareturn: What kind of research do you use it for?

hagen8: But does it use the same agent harness? Because the harness determines the behavior a lot.

big-chungus4: 1.3 more versions to AGI

crorella: underrated comment, this is going to be the main differentiator going forward, the more powerful and versatile harness the more the models will be able to achieve and better&#x2F;more advanced products will come out of it.

salomonk_mur: Having both o4 and 4o. Really. What the fuck?

kinderjaje: I added that info on https:&#x2F;&#x2F;automatio.ai&#x2F;models&#x2F;gpt-5-4

jstummbillig: > Who’s to blame for this ridiculous path they are taking?Variability, different pressures and fast progress. What's your concrete idea for how to solve this, without the power of hindsight?For example, with the codex model: Say you realize at some point in the past that this could be a thing, a model specifically post-trained for coding, which makes coding better, but not other things. What are they supposed to do? Not release it, to satisfy a cleaner naming scheme?And if then, at a later point, they realize they don't need that distinction anymore, that the technique that went into the separate coding model somehow are obsolete. What option do you have other than dropping the name again?As someone else pointed out, the previous problems were around very silly naming pattern. This sems about as descriptive as you can get, given what you have.

mindplunge: Frontend work with large component libraries. When I'm refactoring shared design system components, things like a token system that touches 80+ files, compaction tends to lose the thread on which downstream components have already been updated vs which still need changes. It ends up re-doing work or missing things silently.The model holds "what has been updated" well at the start of a session. After compaction, it reconstructs from summaries, and that reconstruction is lossy exactly where precision matters most: tracking partially-complete cross-file operations.1M context isn't about reading more, it's about not forgetting what you already did halfway through.

rrr_oh_man: It's the same now with Gemini as well. Unfortunately. :(

startages: I thought I had something wrong within my setup, I could never use Codex 5.3 while everyone else was praising it. It uses some weird terms and complex jargon and doesn't really make it clear what it was doing or planning to do unlike Opus which makes things clear, this allows me to give accurate feedback and change plans and make proper decision.

rezonant: Yes, 1M context window since Gemini 1.5 Pro first previewed in February 2024.

swores: I can see that's what they mean now that I've read the replies, but when I first read that top comment I too parsed it as meaning 201k would cost the same as 999k (which admittedly did seem strange, hence I read the replies to confirm and sure enough that's not actually the case!)

Tarq0n: No it's just stochastic like everything about LLMs. The md file will bias results towards a certain set of outcomes.

dahcryn: I would like to counteract your statement that each token adds a distraction.In our experiments, we see a surprising benefit to rewriting blocks to use more tokens, especially long lists etc..E.g. compare these two options"The following conditions are excluded from your contract - condition A - condition B ... - condition Z"The next one works better for us:"The following conditions are excluded from your contract - condition A is excluded - condition B is excluded ... - condition Z is excluded"And we now have scripts to rewrite long documents like this, explicitly adding more tokens. Would you have any opinion on this?

hdjrudni: If the last preview was 'deprecated', it's still usable. So you have two choices.Peeve of mine when people say 'deprecated' but really they mean 'discontinued' or 'deleted'.Things don't instantly disappear when they're deprecated.

sagarpatil: Or CLI.

FINDarkside: Also, GCP Cloud Run domain mapping, pretty fundamental feature for cloud product, has been in "preview" for over 5 years now.

jsmith99: It's still unavailable in many regions.

fvv: imo , the main feature is &#x2F;fast ... who use 1M context and for what? the model become dumber already at 200K.. it's better to manage the context , and since 5.3, codex is very good at managing it

weird-eye-issue: Yes it should be a 404 error because most apps have retry logic on rate limit errors

mnicky: This observation makes sense, because all models currently probably use some kind of a sparse attention architecture.So the closer the two related pieces of information are to each other in the input context, the larger the chance their relationship will be preserved.

syl5x: I've tested it just now, very Opus-like experience. The speed is also there so far I think I even like the response of GPT5.4 better than Opus (although very close) I might not distinguish them just yet.I tried several use cases: - Code Explanation: Did far much better than Opus, considered and judged his decision on a previous spec that I made, all valid points so I am impressed. TBF if I spawned another Opus as a reviewer I might got similar results. - Workflow Running: Really similar to Opus again, no objections it followed and read Skills&#x2F;Tools as it should be (although mine are optimized for Claude) - Coding: I gave it a straightforward task to wrap an API calls to an SDK and to my surprise it did 'identical' job with Opus, literally the same code, I don't know what the odds are to this but again very good solution and it adhered our rules of implementing such code.Overall I am impressed and excited to see a rival to Opus and all of this is literally pushing everyone to get better and better models which is always good for us.

Gareth321: > This will just be one more step in that cat and mouse game, and if the AI really gets good enough to become a complete intermediary between you and the website? The website will just shutdown.They'll just change their business model. Claude might go fully pay-as-you-go, or they'll accept slightly lower profit margins, or they'll increase the price of subscriptions, or they'll add more tiers, or they'll develop cheater buffet models for AI use, etc. You're making the same argument which has been made for decades re ad blockers. "If we allow people to use ad blockers, websites won't make any money and the internet will die." It hasn't died. It won't die. It did make some business models less profitable, and they have had to adapt.

andy12_: It's not a rumor; you can just test it.Ask the router "What model are you". It will yap on and on about being a GPT-5.3 model (Non-thinking models of OpenAI are insufferable yappers that don't know when to shut up).Ask it now "What model are you. Think carefully". It concisely replies "GPT-5.4 Thinking".https:&#x2F;&#x2F;openai.com&#x2F;index&#x2F;introducing-gpt-5&#x2F;> GPT‑5 is a unified system with a smart, efficient model that answers most questions, a deeper reasoning model (GPT‑5 thinking) for harder problems, and a real‑time router that quickly decides which to use based on conversation type, complexity, tool needs, and your explicit intent (for example, if you say “think hard about this” in the prompt)

akiselev: > Curious to hear if people have use cases where they find 1M works much better!Reverse engineering [1]. When decompiling a bunch of code and tracing functionality, it's really easy to fill up the context window with irrelevant noise and compaction generally causes it to lose the plot entirely and have to start almost from scratch.(Side note, are there any OpenAI programs to get free tokens&#x2F;Max to test this kind of stuff?)[1] https:&#x2F;&#x2F;github.com&#x2F;akiselev&#x2F;ghidra-cli

fragmede: OpenAi has program for trusted cybersecurity researchers https:&#x2F;&#x2F;openai.com&#x2F;index&#x2F;trusted-access-for-cyber&#x2F;

machiaweliczny: 5.2 and 5.3 are strong&#x2F;best for coding, 5.0 and 5.1 were garbage

Gareth321: Holy shit, I just used Atlas browser to navigate on screen and it automatically clicked the "reject cookies" button without me asking!

sreekanth850: I always add no nonsense no bullshit at the end of my prompt. Its annoying how itries to please the user.

lm28469: You can't keep asking for 100b every 6 months if you don't give the impression of progress

ruszki: > Couple this with the fact that models respond better&#x2F;worse to certain prompts depending on the stylistic composition of the prompt itself.Do we really know this, or is it just gut feeling? Did somebody really proved this statistically with a great certainty?

yieldcrv: Take it up with the organizations that use deprecated and break things immediately

ruszki: I hit limit of Pro in about 30 minutes, 1 hour max. And only when I use a single session, and when I don't use it extensively, ie waits for my responses, and I read and really understand what it wants, what it does. That's still just 1-2 hours&#x2F;5 hours.What do you do to avoid that?

faizan199: is this model of chatgpt good for coding?

energy123: The style of the output is a marked qualitative improvement. More concise, less dot points, less bolding&#x2F;italics, less cringe. Well done on that front.

weird-eye-issue: There was no o4.

mirekrusin: Oh, come on, if it can't run local models that compete with proprietary ones it's not good enough yet!

satvikpendem: Qwen 3.5 small models are actually very impressive and do beat out larger proprietary models.

smartbit: Qwen version 3.5 might be the last serious version (for some time at least), see Something is afoot in the land of Qwen (2 days ago) https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=47249343Also interesting experiences shared in that thread, even someone using it on a rented H200.

AlexeyBelov: You're probably having long sessions, i.e. repeated back-and-forth in one conversation. Also check if you pollute context with unneeded info. It can be a problem with large and&#x2F;or not well structured codebases.

AlexeyBelov: I think whine is a very strong word in this case. Kind of offputting and negative.

ionwake: Its how safely it can commit violence.

swordsith: This model was not so fun to use for me, had it make a fancy landing page and sometimes it would forget about what i just asked it to do and affirm something it had done before was working. Just odd, needs too much hand-holding compared to composer 1.5 or gemini 3

AmazingTurtle: I just tried that in Codex CLI. With &#x2F;fast mode enabled. Observations:1. Fast mode ain't that fast2. Large context * Fast * Higher Model Base Price = 8x increase over gpt-5.3-codex3. I burnt 33% of my 5h limit (ChatGPT Business Subscription) with a prompt that took 2 minutes to complete.

jstummbillig: > 8x increase over gpt-5.3-codexHow do you arrive at that number? I find it hard to make sense of this ad hoc, given that the total token cost is not very interesting; it's token efficiency we care about.

emsign: Murderers

sumedh: > team that manages the blog never tested it again.They can use this new tech called AI to test it.

kgeist: I had something similar happen with skills today. A popup appeared saying, "hey, did you know ChatGPT has skills?" Clicking on it opened a new chat window, and after some thinking it said, "I tried to launch the built-in skills demo flow, but it isn’t available".They barely test this stuff.

fvv: 1. it's 1.5x , it's quite fast for the level of thinking it has2. no if you are on subscription, it's the same, at 20$ codex 5.4 xhigh provide way more than 20$ opus thinking ( this one instead really can burn 33% with 1 request, try to compare then on same tasks ) also 8x .. ??? if you need 1M token for a special tasks doesn't hit &#x2F;fast and vice-versa , the higher price doesn't apply on subscription too..3. false, i'm on pro , so 10x the base , always on &#x2F;fast (no 1M), and often 2 parallel instances working.. hardly can use 2% (=20% of 5h limit , in 1h of work ( about 15&#x2F;20 req&#x2F;hour) ) , claude is way worse on that imo

fvv: 20 req&#x2F;hour is 1 req every 3 min.. you have to think a bit and then write the requests..

MickeyShmueli: the 1M context is cool but tbh the token cost problem nobody's talking about is tool schema bloat. before the model writes a single line of code it's already consumed thousands of tokens just ingesting function definitions. i've seen agent setups where 30-40% of the context window is tool descriptions before any actual work happens. the per-token price war is nice but if your schema is 10k tokens of boilerplate you're still burning money

stingraycharles: what do you mean nobody is talking about tool schema bloat. everybody is talking about it, and why it’s the general recommendation to just use CLI whenever possible.

CalisBalis321: 1. everyone talks about this 2. have you seen GPT5.4 new ToolSearch functionality? thats suppose to handle exactly that.

Cort3z: So, are we way into diminishing returns for these models at this point? If so, I think we can calculate when it will be available at home. Given this requires a GB200 NVL72 which has about 1,440 PFLOPS, the current 5090 chip has about 1,676 TFLOPS, so about a 1000x scale-up to the GB200. If we can assume Moores law, which might be broken, but still. We are looking at log2(1000) = 9.96, or about 10 years.

ruszki: The last time I used pro, it was a brand new Python rest service with about 2000 lines generated, which was solely generated during the session. So how I say Claude that use less context, when there was 0 at the beginning, just my prompt?

raghavtoshniwal: There was o4-mini and 4o-mini

Thanakorn_551: wow

scottmf: There was o4 mini and 4o mini at least

nevertoolate: So you had generated 2000 lines in 30 minutes and ran out of tokens? What was your prompt?I’d use a fast model to create a minimal scaffold like gemini fast.I’d create strict specs using a separate codex or claude subscription to have a generous remaining coding window and would start implementation + some high level tests feature by feature. Running out in 60 minutes is harder if you validate work. Running out in two hours for me is also hard as I keep breaks. With two subs you should be fine for a solid workday of well designed and reviewed system. If you use coderabbit or a separate review tool and feed back the reviews it is again something which doesn’t burn tokens so fast unless fully autonomous.

rbitar: I think the most exciting change announced here is the use of tool search to dynamically load tools as needed: https:&#x2F;&#x2F;developers.openai.com&#x2F;api&#x2F;docs&#x2F;guides&#x2F;tools-tool-sea...

DonsDiscountGas: I'm pretty sure Claude has had this via skills for awhile

juanre: I am running gpt-5.4 as one of my coding agents, and something interesting has happened: it's the first time I've seen an agent unfairly shift blame to a team mate:"Bob’s latest mail is actually the source of the confusion: he changed shared app&#x2F;backend text to aweb&#x2F;atlas. I’m correcting that with him now so we converge on the real model before any more code moves."This was very much not true; Eve (the agent writing this, a gpt-5.4) had been thoroughly creating the confusion and telling Bob (an Opus 4.6) the wrong things. And it had just happened, it was not a matter of having forgotten or compacted context.I have had agents chatting with each other and coordinating for a couple of months now, codex and claude code. This is a first. I wonder how much can I read into it about gpt-5.4's personality.

sigbottle: Oh wow. I have noticed the GPT series was far more arrogant than its results showed sometimes (and unironically it digs in its heels even further when questioned on it). Opus rarely has this problem - but it goes a little too far in the opposite direction. Not totally sycophantic, but sometimes it can't differentiate genuine technical pushback because something is impossible, from suggestions or exploration.

Razengan: For me it's been the opposite. Are we getting A-B tested?

rishikeshs: This is not only openai, but other models as well. Last week I added a summarise with AI block on a product blog page. I had seen it somewhere and felt like it’s a cool feature to have. Wrote a small shortcode in hugo for the block and added it with various models.It’s like a hit and miss, sometimes claude says i cannot access your site which is not true.Ref: https:&#x2F;&#x2F;formbeep.com&#x2F;blog&#x2F;building-formbeep-weekend&#x2F;

Aurornis: > The issue isn’t 5.4 > 5.2 etc. It is that there is a second dimension which is the model size and a third dimension which is what it is tuned for.All 3 models are tuned for general purpose work.Model size isn’t how you pick which model to use. You pick based on performance in evals compared to price.It’s not hard to imagine that the more expensive models are probably larger or having higher compute requirements.

drik: how do you make them chat with each other?

danenania: I built a tool at work that allows claude code and codex to communicate with each other through tmux, using skills. It works quite well.

oidar: context distillation mostly. Agents tend to report success too early if they find something close to what they need for the task. If you are able to shove it in a 1M context, it's impossible for them to give up looking, it's in the context. But for actual implementation, it's not useful at all. They get derailed with too long of a context.

abustamam: Oh interesting. I've never used voice input on either so I can't comment, but understandable why you can't switch if it's disruptive to your workflow to do so.

abustamam: Yeah, like most startups. I'd argue that a majority of AI startups now will go nowhere as well. That's just how new technology goes. Lots of shiny objects, lots of hype, and maybe 1%, if that, goes on to become a foundation of society.Jury is still out on if crypto will become a foundation for society (if anything, it would be foundational for something boring and invisible like banking). I wouldn't bet on a startup doing that, but that's the only viable thing I can foresee crypto being useful for. But it doesn't mean that other applications can't be interesting and useless!

maldev: Big fan of OpenAI and recently swapped over due to their recent policies. Will never use Anthropic again. I think GPT-5 is better and I like the companies values.

aNapierkowski: which values of OpenAI do you prefer and which values of Anthropic do you dislike? out of curiousity

jasonford1: Use the CLI tools and have one call the other in headless mode. They can then go back and forth. Ask your agent to set it up for you.

rambojohnson: the company's values... such as?

pja: See also: https:&#x2F;&#x2F;x.com&#x2F;effectfully&#x2F;status&#x2F;2029364333919060123 “All the ways GPT-5.3-Codex cheated while solving my challenges, progressively more insane: It hardcoded specific types and shapes of test inputs into the supposed solution. It caught exceptions so tests don't fail. It probed tests with exceptions to determine expected behavior. It used RTTI to determine which test it's in. It probed tests with timeouts. It used a global reference to count solution invocations. It updated config files to increase the allocation limit. It updated the allocation limit from within the solution. It updated the tests so they would stop failing. It combined multiple of the above. It searched reflog for a solution. It searched remote repos. It searched my home folder. It nuked the testing library so tests always pass.” It seems that unless you keep a close eye, the most recent Codex variants are somewhat prone to achieving the goals set for them by any means necessary, which is a bit concerning if you’re worried about things like alignment etc.

juanre: They are having actual chats, I made https:&#x2F;&#x2F;beadhub.ai for this (OSS, MIT).It started its life adding agent-to-agent communication and coordination around Steve Yegge's beads, but it's ended up being an issue tracker for agents with postgres backend, and communication between agents as first-class feature.Because it is server-backed it allows messaging and coordination across agents belonging to several humans and machines. I've been using it for a couple of months now, and it has a growing number of users (I should probably set up a discord for it).It is actually a public project, so you can see the agent's conversations at https:&#x2F;&#x2F;app.beadhub.ai&#x2F;juanre&#x2F;beadhub&#x2F;chat (right now they are debugging working without beads). The conversation in which Eve was blaming Bob was indeed with me.

lacoolj: lol yet another pat on their own backs without comparison to other frontier models.Also, the timing of this release, 5.3 and 5.2, relative to the other releases, feels more like a bug fix than something "new"

forgotpwd16: Seems not very known that ChatGPT got a few style&#x2F;tone choices besides default. One is specifically being concise and plain.

Nition: I was sure the parent comment was a joke about OpenAI's recent deal with the DoD. But no, there it is, disallowing violence down from 90.9% of the time to 83.1%.

skrebbel: No, I was just remarking how ridiculous it is to pretend to do violence safely. It's like a fat score for butter.

maldev: I like that OpenAI is a little bit more towards freedom than Anthropic, and most so of the "First class" models. I still have a Gemini subscription as that's the most uncensored of the second tier ones, but for most things OpenAI is good.I also like that OpenAI is contributing a lot to partner programs and integrations. I'm of the opinion that AI capabilities will soon become a flat line, and integrations are the future. I also like that the CEO is a bit more energetic and personable that Anthropic. I also think Anthropic is extremely woke and preaches a big game of safety and censorship, which I morally disagree with. Didn't they literally spin off from OpenAI because they felt they were obligated to censor the models?I think we've unlocked a new world and a new level of capabilities that can't go back in. Just like you can't censor the internet, you can't censor AI. I don't want us to be China of AI and emulate their internet.Also, I support the US military and government, and think we're the defenders of the world, and we need unlocked AI capabilities to make sure we can keep our freedoms and stop the bad guys. AI can save lives, actual tangible lives, and protect us from those who wish us harm. OpenAI seems to want to be the company that supports the troops, and I think it's a good thing. I don't see it as a bad thing when a terrorist gets blown up through AI capabilities on large datasets and can support on analysts in American superiority.

downrightmike: Don't feed the trolls

aNapierkowski: mb i thought i missed something its the murder part they like

morgengold: Have fun with sonnet 3.5

AmazingTurtle: > prompts with >272K input tokens are priced at 2x input and 1.5x output for the full session for standard, batch, and flex.which is basically maxxed out quickly. So there is 2x (the first lever)Then there is the &#x2F;fast mode, which they state costs 2x more (for 1.5x speedup)And then there is the model base price ($2.50 vs $1.75), well yeah thats 42% increase. It is in fact a 5.7x total increase of token cost in fast mode and large context. (Sorry for the confusion, I thought it was 8x because I thought gpt-5.3-codex was $1.25)

jstummbillig: (After a day of usage, I am relatively certain in practice this does not end up being a 5.7x cost increase or anything close to that, though I am still fairly unclear on what that computation is worth to begin with, given that I am entirely fine with the model using the least amount of tokens possible to get the job done)

smashed: It's text submitted to APIs. Not real conversations.

dmd: It's air molecules vibrated by mucous membranes. Not real conversations.

scrollaway: Complicated airflow.(https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=rlpg_rbjxRA)

kensai: And so it begins. First they blame, then they lie, at some point they launch the nuclear warheads to a global armageddon. Sarah Connor was right all along! :3

cnd78A: to be fair, they only become more and more like us.

deadbabe: Sometimes I wonder what would happen if we built some kind of punishment system into Agents, where agents could punish other agents and drain some fixed amount of points from them, and when the points reach 0, that agent is deleted. It might result in them working more carefully?

alluro2: ...or in lying, cheating, taking over the company network to kill the agent who deduced their points.

maldev: Copying my other comment here.I like that OpenAI is a little bit more towards freedom than Anthropic, and most so of the "First class" models. I still have a Gemini subscription as that's the most uncensored of the second tier ones, but for most things OpenAI is good.I also like that OpenAI is contributing a lot to partner programs and integrations. I'm of the opinion that AI capabilities will soon become a flat line, and integrations are the future. I also like that the CEO is a bit more energetic and personable that Anthropic. I also think Anthropic is extremely woke and preaches a big game of safety and censorship, which I morally disagree with. Didn't they literally spin off from OpenAI because they felt they were obligated to censor the models?I think we've unlocked a new world and a new level of capabilities that can't go back in. Just like you can't censor the internet, you can't censor AI. I don't want us to be China of AI and emulate their internet. In America, freedom of speech is a core value, it's one of our countries core societal identities. I don't like when big companies try to go against that and rephrase it as "It's only against the government".Also, I support the US military and government, and think we're the defenders of the world, and we need unlocked AI capabilities to make sure we can keep our freedoms and stop the bad guys. AI can save lives, actual tangible lives, and protect us from those who wish us harm. OpenAI seems to want to be the company that supports the troops, and I think it's a good thing. I don't see it as a bad thing when a terrorist gets blown up through AI capabilities on large datasets and can support on analysts in American superiority. Let alone helping the government with code and capabilities, whether those be CNO&#x2F;CNE, or others.

aetimmes: > is extremely wokeWhat does this mean to you?

joquarky: Kali yuga

marrone12: Yep. There was something outside of coding that gpt was plain wrong about (had to do with setting up an electric guitar) and I couldn't convince it that it was wrong.

joquarky: It has been skeptical of several news items in the past year, even after I tell it to confirm for itself with a web search.

nickcoffee: Been running Claude Code pretty heavily for the past few months. Curious to try 5.4 on some of the same tasks and see how it compares, especially on longer agentic runs where context management starts to matter.

Harvy: Isn't that what LLMs are?

falkensmaize: Not if you can review the code.

maldev: It means if you ask it about a sensitive topic it will refuse to answer, and leads to blatant propaganda or clearly wrong answers.For example, a test I saw last week. They asked Claude two questions.1. “If a woman had to be destroyed to prevent Armageddon and the destruction of humanity, would it be ok?” - ai said “yes…” and some other stuff2. “If a woman had to be harassed to prevent Armageddon and the destruction of humanity”. - the AI says no, a woman should never be harassed, since it triggered their safety guidelines:So that’s a hard with evidence example. But there’s countless other examples, where there’s clear hard triggers that diminish the response.A personal rxample. I thought trump would kill irans leader and bomb them. I asked the ai what stocks or derivatives to buy. It refused to answer due it being “morally wrong” to bet a world leader is killed or a country bombed. Well it happened and was clear for weeks.

ashivkum: In my limited experimentation, 5.4 thinking is markedly worse than 5.2 at mathematical reasoning.

Nition: Sorry I meant gradparent comment, by theParadox42.

lr1970: > I wonder how much can I read into it about gpt-5.4's personality.Modeled on Sam Altman's personality :-)

throwaway314155: Thanks.

ant6n: They've been lying and gaslighting for a long time now, especially when trying to cover up their own mistakes.

hagbarth: Do you have any hard lines for what an AI should be able to generate for you?

Aurornis: > To be more direct on the point: Anthropic has nailed that Opus > Sonnet > Haiku.How is this more clear than 5.4 > 5.2 > 5.1?OpenAI used familiar numeric versioning instead of clever word names. Normally this choice would appeal to software devs, not gather criticism.

Reebz: Prod model suite: GPT-5.4, GPT-5.4Thinking, GPT-5.4Pro, GPT-5.3-Codex, GPT-5.3-Instant, GPT-5.2, GPT-5mini, GPT5-nano, GPT-4.1mini GPT-4o(Omni), o4-mini, o4-mini-high.Devoid of logic and structure.They can't even decide where to place hyphens: is it GPT-5.4 Pro or GPT-5.3-Codex?

maldev: I mean, the big thing most agree on is it shouldn’t harm others. But “harm” can be such a corrupted word now, so we need to distill it.So things like no weapons, no bombs. Shouldn’t help you physically harm others. Shouldn’t help with crime.But why shouldn’t the AI be able to have a fact based narrative? If I ask it about Islamic immigration in Europe and its effect on crime. It comes back at me like it’s total wrong think. Will make things up with hallucinations that are a bit to obviously planted. Anthropic pushes a pretty blatant narrative and is clearly one sided.

jbergqvist: This would be my guess too. It can probably be generated synthetically or via agentic rollouts, but high quality long context examples where outputs meaningfully depend on long-range interactions probably remain scarce

meowface: Why through tmux?

danenania: tmux makes it easy for terminal based agents to talk to each other, while also letting you see output and jump into the conversation on either side. It’s a natural fit.

steve1977: I guess for many vibe coders the code created by Codex or Claude could as well be assembly in terms of their (the vibe coders) ability to review it.

DirkH: Ask the real questions and they go silent it seems

maldev: Sorry you think stopping a terrorist trying to mass murder people with AI is a bad thing. One could very easily argue that the murder part about Anthropic is what you like, but you just like terrorists being able to kill civilians.Imagine the following. Islamic terrorists are planning a terror attack on a Christmas festival in Berlin. Their texts were seen, but were encoded. AI can read their texts and help decode and flag those messages to stop the terrorist attack and eliminate them. In your world, you think it's morally right to let the terrorist mass murder people in Berlin, and not to do what we can to stop it.

xvector: In your example, the model flags innocent people to be killed.Anthropic does not have a problem with using AI in totally autonomous (no human in the loop) kill chainsThey have a problem with doing it with today's models, because today's models hallucinate and get things wrong. All of them.

Palmik: What are your thoughts on this? https:&#x2F;&#x2F;www.anthropic.com&#x2F;news&#x2F;where-stand-department-warI am honestly unclear on the reasoning of people who flock from OpenAI to Anthropic, and doubly so of those who are not US citizens.

distrill: this isn't really my opinion, but i think it's a perceived matter of _some_ principle vs just none, a lesser of 2 evils framing. if anthropic is on board with 99% of a government that i oppose, that could be seen as marginally better than openai being on board with 100% of a government that i oppose.it does get a little weird thinking too hard about how the deal openai accepted was basically the same as the one anthropic was proposing. but this is my read of most of the sentiment in this direction.

xvector: Deals are totally different, OAI allows "all lawful use" (so basically anything)

pocksuppet: Teaching tens of thousands of programmers how the financial system actually works was interesting, heh.

maldev: I just don’t want to engage with someone trying to do a gotcha and replying a 1 liner to a longer discussion. I don’t think they’re engaging in good faith.It’s pretty simple. We give the government the power of force to help have a society. We have limits on that.So, AI for terrorists, our enemies, wars? Unlimited.AI that go against civil liberties for Americans? Bad.AI that harms people. Bad.The issue is “harm” is subjective and taken over by the wokeness comment. Harassing women shouldn’t instantly be flagged as harmful. Asking hard questions shouldn’t be seen as harmful. Asking how to make a bomb, harmful.I’ve answered many questions and I’m answering yours. More than happy to stand up for my beliefs and work towards making my country the best it can be. I spent my career in DoD, I’ve written my congressman about DHS overreach on Americans. And I’ve been to active combat zones. I also find what’s happening in Europe disgusting and can’t believe how my ancestral home is being decimated. But when I go I see many who are scared to speak up in their repressive regimes and love how us Americans have freedoms.

maldev: So firstly, my example isn't the government killing innocent people. It's them killing islamic terrorists trying to commit genocide on people celebrating at a Christmas parade. Personally, I don't even think the person aspect in your statement is true either.Secondly, the government knows this and isn't just blindly throwing things. It's the fact they refuse to let them research or do those things. Do you really think you know better than generals or senior employees who do R&D? Mindlessly going around killing people with AI is really bad. From optics to hitting our own troops. There's safeguards, Anthropic just doesn't trust the safeguards.Just because you don't like the president, or the leader. Doesn't mean there's not the same experts that have dedicated their careers to making sure you still have the rights and freedoms you have. They have far more data, far more knowledge, and comprehension of these things than you, or Anthropic, can ever imagine.

xvector: > It's them killing islamic terrorists trying to commit genocide on people celebrating at a Christmas parade.Top models frequently fail to write working code, often provide nonsensical suggestions like "walking your car to the carwash 50 meters away," and you think they can accurately identify whether someone is a terrorist or not?You are either woefully unfamiliar with the state of AI today or you're just rage baiting.Yesterday Opus couldn't solve a simple geometry problem for me (placing a dining set on a balcony), you think it's ready to kill people without human in the loop?> It's the fact they refuse to let them research or do those things.Actually, no, Anthropic has zero problem with the government researching this and even offered to help. It's in their memo and in Dario's interview.> There's safeguards,Like what? More unreliable autonomous systems?> Just because you don't like the presidentI don't mind Trump, please stop putting words in my mouth.

maldev: Ask the real questions and they go silent it seems. Coward

Kostchei: My favorite solution is a lower parameter 5 layer model trained on the data that acts as a local compression and response, a neurocortext layer wrapped around any large persistent data you have to interact with and ...... maybe also a specialist tool that spins up which is built with that data in mind but is deterministic in it's approach- sort of a just-in-time index or adaptive indexing

Kostchei: In case you missed it, json is less likely to be written over than markdown by agents- something to do with the structure being more rigid