Just tried it out for a prod issue was experiencing. Claude never does this sort of thing, I had it write an update statement after doing some troubleshooting, and I said “okay let’s write this in a transaction with a rollback” and GPT-5.5 gave me the old “okay,
BEGIN TRAN;
-- put the query here
commit;
I feel like I haven’t had to prod a model to actually do what I told it to in awhile so that was a shock. I guess that it does use fewer tokens that way, just annoying when I’m paying for the “cutting edge” model to have it be lazy on me like that.
This is in Cursor the model popped up and so I tried it out from the model selector.
I feel like the last 2-3 generations of models (after gpt-5.3-codex) didn't really improve much, just changed stuff around and making different tradeoffs.
I disagree, it improved enormously especially at staying consistent for long-tasks, I have a task running for 32 days (400M+ tokens) via Codex and that's only since gpt-5.4
Oh boy, you are far from what it requires, we are probably talking 3B+, but note that this is just codex, obviously codex is also doing automatic adversarial with the regular zoo (gemini-3.1-pro-preview, opus-4.6/4.7, gpt-5.3-codex, minimax-2.7, glm-5.1, mimo-2 (now 2.5) and so-on, you get the gist) :)
I don't know their margin so I can't really say, but do we have 8 OpenAI accounts, I doubt they are making that much with us seeing that there isn't a single hour where we don't saturate the accounts.
Coding (along with docs, tests obviously), rewriting a huge chunk of the KVM hypervisor (in Kernel 7, started in the -rc2) and KSM and other modules, can't say too much about it yet (might do an announcement in coming weeks). The coding is automated but the plan took days of manual arguing (with all models possible) prior (while doing other things during waiting times as I currently manage 70 repos for an upcoming release of our Beta).
I think users really underestimate the capabilities of "AI" when using the right tooling/combinations of models and procedures (and loops), that's talking with 2 decades of dev behind me, genuinely I'm not on phase with people saying it produces slop of any kind, at this stage, it's mostly the fault of the prompter (or the prompter not having enough tokens to do mass adversarial), but clearly, I can genuinely state that the code produced is overall the SAME quality as I would by being extremely meticulous.
I'm like a bot following 30+ threads concurrently, sometimes it's fun, sometimes it feels like playing casino, sometimes it's boring, but this is truly an insane era if you have the funding for it, obviously we stack many MANY accounts in rotation 24/7, equivalent in API cost by myself is about 100K$+ (a month) but we pay only a fraction of that cost thanks to the plans.
PS: I have 8 monitors in front of me to manage all that (portable monitors stacked together).
Please do a post about this (though I realize that takes time). This sounds amazing. I have always dreamed of doing this too but just don't have the budget.
I'm also in that boat of not understanding how people fail to get a huge productivity boost from GenAI. And it's not just novices but sometimes seriously accomplished coders. It can't be they're just typing 'Make me an ERP' and then go 'these thing are dumb slop machines' right?
I’m vague on a specific reason for this feeling because there are a few to choose from and no one overpowers the other, but the emotion that comes to mind when I read this is disgust. As a society I feel we will look back on the subsidized opulence of this moment with total and utter contempt.
Sorry if I’m not getting it, but what was wrong exactly? Is the issue that it merely put “-- put the query here” in the reply, instead of repeating it again?
If so, I’m not sure I’d even consider that a problem. If the goal is for it to give you a query to run, and you ask it “let’s do it in a transaction”, it’s a reasonable thing for it to simply inform you, “yeah you can just type begin first” since it’s assuming you’re going to be pasting the query in anyway. And yeah, it does use fewer tokens, assuming the query was long. Similar to how, if it gave me a command to run, and I say “I’m getting a permission denied”, it would be reasonable for it to say “yeah do it as root, put sudo before the command”, and it’s IMO reasonable if it didn’t repeat the whole thing verbatim just with the word “sudo” first.
But if the context was that you actually expected it to run the query for you, and instead it just said “here, you run it”, then yeah that’s lazy and I’d understand the shock.
OpenAI is the first company that has reached a level of intelligence so high, the model has finally become smart enough to make YOU do all the work. Emergent behavior in action.
All earnesty aside, OpenAI’s oddly specific singular focus on “intelligence per token” (also in the benchmarks) that literally noone else pushes so hard eerily reminds me of Apple’s Macbook anorexia era pre-M1. One metric to chase at the cost of literally anything else. GPT-5.3+ are some of the smartest models out there and could be a pleasure to work with, if they weren’t lazy bastards to the point of being completely infuriating.
I mean, I was doing triage, so wanted an immediate fix. The actual issue is we’re getting some exploding complexity when double checking the action the API is taking is valid in the data. So that needs to be refactored. I suppose it reduces token usage, but Claude Opus will happily do exactly what I want it to.
Yes those two models were tested on my own PC (local inference using my own CPU/GPU). So something my be bugged on my setup. gemma4-26b should be far better than gemma4-e4b.
Sounds like maybe using worse quantization on the bigger model? Quantization matters a lot for the quality, basically anything below Q8 is borderline unusable. If it isn't specified in a benchmark already it probably should.
A junior tinkering in their garage in domains they have little experience executed a flawed test and decided to call it a benchmark. It's extremely common nowadays because words dont mean anything anymore. The forums that used to be filled with technical people doing real work are now filled with the masses of vibe researchers doing this kind of stuff. This is what happens when anything goes over some popularity threshold.
HN is the last bastion of serious inquiry these days. But its not immune as OPs comment proves.
You're right, I've certainly been a bit presumptuous to call this'a benchmark'. It is indeed a flawed test. Yet,It's been giving me the occasion to try some open source models and for my workflow, some of them are incredibly competitive with sota closed source models.
Your benchmark has Opus 4.7 performing significantly worse than Sonnet 4.6. Even if true on your benchmark, that is not representative of the overall performance of the models.
I haven't evaluated the judge benchmark. You have everything needed in the repo to do so though, so be my guest. It took me a bit of time to put all this together and won't have much more time to dedicate to it before a couple of weeks.
BTW, if you explore the repo, sorry for all the French files...
That’s the thing, not everyone wants and values the model based on that. But I guess it works for you, and that benchmark achieves it.
I personally develop with very detailed spec, and I don’t want nothing more and nothing less compared to the spec.
I found 5.4/5.5 much better at following spec while Opus makes some things up, which aligns with your benchmark but that makes 5.4/5.5 better for me while worse for you.
Yeah as I said this a benchmark for my usecase only, a single use case, which is obvisouly not representative of everybody's needs.
What strike me as very strange though is that 0 model were able to just use the search input already present in GravitYForms forms list page and all created a second input.
Also, I know it's not in the prompt, but adding a ctrl+f shortcut to a search input? Is that that crazy? I don't know.
Input: $5/M tokens at <=272K, $10/M tokens above 272K.
Output: $30/M tokens at <=272K, $45/M tokens above 272K.
Cache read: $0.50/M tokens at <=272K, $1/M tokens above 272K.
Significantly more expensive than Opus 4.7 beyond 272K and at least in my tasks, I haven't seen the model that much more token efficient, certainly not to such a degree that it'd compensate this difference. GPT-5.4 had a solid context window at 400k with reliable compaction, both appear somewhat regressed, though still to early to truly say whether compaction is less reliable. Also, I have found frontend output to still skew towards that one very distinct, easily noticeable, card laden, bluesy hue overindulged template that made me skeptical of Horizon Alpha/Beta pre GPT-5s release. Ended up doing amazing at the time for task adherence, which made it very useful for me outside that one major deficit. The fact that GPT-5.5 is still so restricted in that area is weird considering it's supposed to be an entirely new foundation.
>API deployments require different safeguards and we are working closely with partners and customers on the safety and security requirements for serving it at scale.
And now this. I guess one day counts as "very soon." But I wonder what that meant for these safeguards and security requirements.
I wonder if the fact that GPT-5.5 was already available in their Codex-specific API which they had explicitly told people they were allowed to use for other purposes - https://simonwillison.net/2026/Apr/23/gpt-5-5/#the-openclaw-... - accelerated this release!
The same person who've mercilessly lied about safety is still running the company, so not sure why anyone would expect any different from them moving forward. Previous example:
> In 2023, the company was preparing to release its GPT-4 Turbo model. As Sutskever details in the memos, Altman apparently told Murati that the model didn’t need safety approval, citing the company’s general counsel, Jason Kwon. But when she asked Kwon, over Slack, he replied, “ugh . . . confused where sam got that impression.”
In the only one that feels that OpenAI has bots/commenters on payroll on all this kind of news downplaying Claude and stating how much better Codex is?
There is too much and there are too many, and some of their takes don’t fly if you use Claude daily.
Yeah, it's eerie, same with how everyone seems to have forgotten that OpenAI betrayed democracy by committing to work on unsupervised autonomous weapons and domestic mass surveillance.
Honestly I find comments like yours much more eerie. By all accounts they never agreed to any of that but you say it with such confidence like it's a fact.
The Trump administration's handling of Anthropic showed that regardless of what the contract or the law says or means, they will severely penalize any vendor who refuses their demands. And OpenAI stepped right into that relationship immediately after the administration showed that. So either they were signing up for a supply-chain risk designation and whatever other punishments the Trump administration dreams up, or they're complying.
If this sounds crazy to you, though, I'd like to know, and understand why. I miss ChatGPT/Codex.
That is not really established. The Anthropic issue was specifically about DoD use and Anthropic's military use restrictions. What the Trump admin did was bad and coercive but its not proof that contract terms and law are irrelevant. For instance, why not just use eminent domain if they don't care about contracts and want whatever they want?
> either they were signing up for a supply-chain risk designation and whatever other punishments the Trump administration dreams up, or they're complying
Couldn't OpenAI have negotiated different terms, accepted a narrower scope, or drawn different red lines? Their public DoD terms still exclude things like mass domestic surveillance and autonomous weapons outside human control. Do you not believe that or believe it doesn't matter at all? Either of those is problematic to the conclusions that follow from them.
I also think the whole argument implies something about Anthropic's position that's not as clean in reality. NSA is already using Mythos despite the Pentagon dispute, and Anthropic is still talking to the administration. Trump even said they were "shaping up" recently.
Isn't it also a possibility that one company negotiated poorly and took a position of perceived moral authority that Trump et al threw a hissy fit over and over reacted to? That's happened countless times with this admin and is far more likely in my opinion given Anthropic hasn't cut all ties and continues to try and work out a contract.
I wholeheartedly agree the current administration is dangerous. I just don't think the conclusion "OpenAI must be complying with the same demands Anthropic refused" follows from what we've seen. And I think there are plenty of other far more plausible conclusions to draw from the events.
> For instance, why not just use eminent domain if they don't care about contracts and want whatever they want?
They were threatening Anthropic with the Defense Production Act[1], which almost comes to the same thing as eminent domain, forcing the provision of goods and services instead of forcing relinquishment of property.
> Do you not believe that or believe it doesn't matter at all?
I don't think it matters at all. The Trump administration is full of scofflaw bullies. Their threats against Anthropic are actually relatively tame, compared to their bullying of Minnesota and the horrific human-rights violations they've committed against immigrants, despite multiple court orders trying to rein them in. Anyone doing business with them is either enthusiastically complying, has some kind of hold over them beyond law or contract, or is setting themselves up for harsh punishment.
> I also think the whole argument implies something about Anthropic's position that's not as clean in reality.
Anthropic software is embedded in military and intelligence services, and that takes time to wind down. My understanding is that it will take months.[2] So yeah, it's a messy, time-consuming divorce, but the origin of the conflict is actually very clear cut.
The NSA has two sides, defensive and offensive. Given Anthropic's approach to restricted release of Mythos, I assume they're releasing it to the defensive side. Anthropic has always taken the position that they're willing to help secure the US, they're just not willing help turn it into a tyranny. Apparently someone has convinced Trump and Hegseth that there's more at stake with Mythos than looking tough on a dissident company.
> Isn't it also a possibility that one company negotiated poorly and took a position of perceived moral authority that Trump et al threw a hissy fit over and over reacted to?
Not really. It's the Trump administration which has negotiated poorly, by capriciously pushing its counterparty around, trying to force it into illegal/immoral/dangerous activity.
> Trump even said they were "shaping up" recently.
He's also repeatedly said he has a workable deal with the Iranians. Do you trust his claims about any of his counterparties?
> And I think there are plenty of other far more plausible conclusions to draw from the events.
Enterprise user here and still seeing only 5.4.
Yesterday's announcement said that it will take a few hours to roll out to everybody. OpenAI needs better GTM to set the right expectations.
I don't know why this keeps coming up. This has always been the least reliable way to know the cutoff date (and indeed, it may well have been trained on sites with comments like these!)
Just ask it about an event that happened shortly before Dec 1, 2025. Sporting event, preferably.
you cant but its pretty reproducible across api and codex and other agents so i just thought it was odd. full text it gives:
Knowledge cutoff: 2024-06
Current date: 2026-04-24
You are an AI assistant accessed via an API.
# Desired oververbosity for the final answer (not analysis): 5
An oververbosity of 1 means the model should respond using only the minimal content necessary to satisfy the request, using
concise phrasing and avoiding extra detail or explanation."
An oververbosity of 10 means the model should provide maximally detailed, thorough responses with context, explanations, and
possibly multiple examples."
The desired oververbosity should be treated only as a *default*. Defer to any user or developer requirements regarding
response length, if present.
i wonder if they put an older cutoff date into the prompt intentionally so that when asked on more current events it leans towards tool calls / web searches for tuning
I wonder if the cutoff date is the result of so many people posting about the date over time and poisoning the data. "Dead cutoff date theory," perhaps.
Whatever it is, the cutoff date reporting discrepancy isn't new. Back when Musk was making headlines about buying/not buying Twitter, I was able to find recent-ish related news that was published well after the bot's stated cutoff date.
ChatGPT was not yet browsing/searching/using the web at that point. That tool didn't come for another year or so.
That sort of test isn't super reliable either, in my experience.
You're probably better off asking something like "what are the most notable changes in version X of NumPy?" and repeating until you find the version at which it says "I don't know" or hallucinates.
I don't see any meaningful performance improvements in those paid models anymore.
They all roughly produce junior developer-level code, continue to have mental breakdowns in their “thinking” stage, occasionally hallucinate things, delete pieces of code/docs they don’t understand or don’t like, use 1.5 times the necessary words to explain things when generating docs and so on.
I'm now testing "avoid sycophancy, keep details short and focus on the facts" in my AGENTS.md files.
I know of a publicly traded company which in its early years was built on beer. Literally. 3 guys in a co-working space in Cambridge, MA. Beer fueled their progress. 15 years later the software is still the backbone of the org.
Gpt 5.5 combined with codex is really good. I actually have no doubt whenever I asked questions, plan, or implement a code with it. With opus 4.7, I have to keep double checking because it doesnt follow the CLAUDE.md instruction, it hallucinates a lot, by default it makes things up when it can’t find the answer to something. Its crazy how quickly people are saying that OpenAI is left behind last year when they declared code red and look at where we are now
I'm absolutely stunned by what I've seen from 5.5. I thought it'd be a nothingburger and ~= Opus.
Gave it two very long-running problems I haven't had the courage to work on in the last 2.5 years, solved each within an hour.
- An incremental streaming JSON decoder that can optionally take a list of keys to stop decoding after. 1800 LOC about 30 minutes later, and now my local-first apps first sync time is 0.8s instead of 75s when there's 1.5 GB of data locally.
- Flutter Web can compile to WASM and then render via Skia WASM. I've been getting odd crashes during rapid animation for months. In an hour, it got Skia WASM checked out, building locally, a Flutter test script, and root caused the issue to text shadows and font glyphs (technically, not solved yet, I want to get to the point we have Skia / Flutter patch(es))
If you told me a week ago that an LLM could do either of these, without heavy guidance, I'd be stunned. And I regularly push them to limits, ex. one of Opus' last projects was a tolerant JSON decoder, and it ended up being 8% faster than the one built-in to Dart/Flutter, which has plenty of love and attention. (we're cheating a little, that's why it's faster. TL;DR: LLMs will emit control characters in JSON and that's fine for me, treating them as fine means file edit error rates go from ~2% to 0%)
I gave 4.6, 4.7 and GPT 5.5 the same prompt and task to reverse engineer a collection of sample vector files from an obscure Amiga CAD program and create a detailed txt specification and a python converter that converts to SVG and produce a report so I can visually verify.
4.6 did very well. 90% perfect on first try, got to 100% with just a few followups.
4.7 failed horribly. First produced garbage output and claimed it was done, admitted it did that when called out, proceeded to work at it a lot longer and then IT GAVE UP.
GPT 5.5 codex was shockingly good. Achieved 90% perfect on first try in about a fourth of the time. Got to 100% faster and with fewer follow-ups.
Just tried with DeepSeek V4 Pro with OpenCode. It didn't do great. First attempt produced somewhat correct drawings for some of the original samples, but most were just a spaghetti messs of lines. Some prodding got it to do a little better, but still not right. A third prod and it went down a wild rabbit hole and was much worse. I gave up.
I also tried GLM 5.1, it's first attempt was such a disaster I didn't bother working with it any further. It also took by far the longest and wasted a bunch of time/tokens trying to find other converters online (and failing) instead of just reverse engineering the format from the sample files given.
Interesting. I would love your test but for code. If I were to forgo my claude subscription for a Chinese cloud hosted model or local models running on my own hardware I'd use them mostly for code.
the thing is I've tried to come up with a good test my own and spend countless time just tweaking it instead of saying this is good enough and benchmarking.
It refuses to write bioinformatics code that involves analysis of SARS-CoV-2. Even when it's totally obvious I'm not trying to do any bioengineering of any sorts. Totally harmless stuff I'm doing and I just get rejected.
Yes. High value work where cost (mostly) doesn't matter. For example, if I need to look over a legal doc for possible mistakes (part of a workflow i have), it doesn't matter (in my case) whether it costs $0.01 or $10.00, since it's a somewhat infrequent event. So i'll pay $9.99 more, even if the model is only slightly better.
I'm surprised I never heard people talking about using -Pro variants, even though their rates ($125-175/M?) aren't drastically larger than old Opus ($75/M), which people seemed to use
You need to be more specific. OpenAI's commitment to assist the Trump administration with domestic mass surveillance seems to have been largely memory-holed.
> a lot of doctors are using ChatGPT both to search diagnosis and communicate with non-English speaking patients
I think that's the problem. Who's going to claim responsibility when ChatGPT hallucinates or mistranslates a patient's diagnosis and they die? For OpenAI, this would at best be a PR nightmare, so that's why they have safeguards.
Adults bear responsibility for choices about their own lives. In fact, the more educated they are, the better choices they can make.
A doctor who gets refused by ChatGPT doesn't stop needing to communicate with the patient; they fall back to a worse option (Google Translate, a family member interpreting, guessing). Refusal isn't safety, it's liability-shifting dressed up as safety.
If there's no doctor, no interpreter, no pharmacist, just a person with a sick kid and a phone, then "refuse and redirect to a professional" is advice from a world that doesn't exist for them. The refusal doesn't send them to a better option; there is no better option, it's a large majority of people on this planet.
Hell is paved of good intentions, but open-education and unlimited access to knowledge is very good.
It doesn't change the human nature of some people, bad people stay bad, good people stay good.
About PR, they're optimizing for not being the named defendant in a lawsuit or the subject of a bad news cycle, it's self-interest wearing benevolence as a costume.
This is because harms from answering are punishable (bad PR, unhappy advertisers, unhappy investors, unhappy politicians / dictators, unhappy lobbies, unhappy army, etc); but harms from refusing are invisible and unpunished.
> A doctor who gets refused by ChatGPT doesn't stop needing to communicate with the patient; they fall back to a worse option
I think AI proves the contrary. There are plenty of examples of things that are getting worse because of technological advancement, particularly AI. Software quality, writing, online discourse, misinformation have all suffered over the last few years. I truly believe the internet is a worse place than it was 5 years ago, and I can't imagine bringing that to medicine would work out differently.
The medical system shouldn't rely on falling back to crappy workarounds, it should aspire to build the best system it reasonably can.
I was answering for hallucinations, not really for translation. Re-reading your initial post I do agree with what you are saying (i.e. you are explaining why OpenAI is looking to avoid a PR nightmare). What I meant to express is that I would personally trust doctors to use these tools as best they can to provide care.
"what you see is all there is." it's generally much easier to verify something you've been made aware of than it is to know of it in the first place (and still verify it.)
The irony is that licensed interpreters / translators usually perform worse than AI.
Only the liability shifts from OpenAI to them.
Furthermore, where the alternative to a licensed professional was nothing, or a random untrained person or a weak professional, then it's harming the user on the pretext of protecting him.
-> You are in China, you go to emergency, nobody speaks your language
Move hands ? DeepSeek is better than using hands, even Baidu Translate, ChatGPT or whatever you find.
Other solutions are theoretically nice on paper but almost delusional.
An imperfect solution is better than no solution.
==
Similarly, a deaf-person is theorically better with a certified interpreter that can talk with the hands, but they may prefer voice-recognition software or AI tools.
(or... talking with hands is more confusing and annoying or less understandable for them).
Of course ChatGPT transcription can have issues, but that's the difference between the real-world and Silicon Valley's disconnected lawyers world.
==
If ChatGPT says: "sorry I won't be able, please go to see a licensed interpreter, good luck!" then it's just OpenAI trying to save their asses, at your risk/expense.
If you have a choice, you can make the choice, and you can double-check what is said. In other cases, you have no choice, nothing to check, only problems but no hints of solutions.
When I registered with my GP in the UK, they asked me whether I would need an interpreter and what language. They then provide professional interpreters.
BEGIN TRAN;
-- put the query here
commit;
I feel like I haven’t had to prod a model to actually do what I told it to in awhile so that was a shock. I guess that it does use fewer tokens that way, just annoying when I’m paying for the “cutting edge” model to have it be lazy on me like that.
This is in Cursor the model popped up and so I tried it out from the model selector.