> Claude Opus 5.5 is our first release since we called for pacing the frontier.
Interesting how the very first line is used to remind the reader of their call to pace the frontier just last week, and everything else after that line is to demonstrate with very specific numbers how they absolutely are not pacing.
I really don't think it's productive for internet forums to constantly be criticizing language choice when the meaning is clear. Better to respond to the substance of the issue than word choice.
Edit: In response to the initial replies. To me it clearly means "releasing frontier models at any pace less than as fast as possible". It implies relative restraint compared to the previous state and without stating the degree of restraint.
The meaning isn't clear at all. So open for interpretation that it is meaningless. That's the whole fucking point. For all I know they are "pacing the frontier", or not. The fact that there's no meaning to it let's you know that it was a pointless waste of tokens and attention.
There's a whole blog article written by Dario Amodei, literally linked from the text "pacing the frontier" - click that and read if you need more context and understanding. The first sentence in this press release is merely stating a fact about the timeline - this is Anthropic's first release since Amodei released the referenced article.
The blog post provides some context but no understanding.
It's meaningless because it's unverifiable. Dario all but said we wouldn't notice if they were pacing or not, because they have no intention of stopping development. The bits we could in theory verify are the external audits, whose independence has already been called into question.
At the very least, dumping a new model on the world before the ink on the glossy brochures of "Pacing the Frontier" was even dry calls into question their commitment.
... in fact the "pace" seems to have all but increased AND the surface area of the "frontier" has all but increased: There's the Pareto frontier, the AGI frontier, the price frontier, the structured output frontier, the open weight frontier, the Chinese chip-trained frontier, the inference frontier, harnesses ...
The only concrete action item in the blog post is a call to increase restrictions exports of GPUs and chip-making equipment to China.
If Amodei were serious, he'd be calling for a ban on investment in AI research. It reminds me of all the "land acknowledgements" by people who have zero intention of giving the land back.
Yes and it was nonsense. Amodei is free to slow up frontier development and focus more on safety and testing any day.
The reason why he and is peers are calling for it to be implemented by somebody else (a legal framework), is for their own financial benefit and to keep competitors out.
What do you think would be the benefit of them stopping if others race ahead? Do you think they have some special ability no one else will find, among the many competitive firms right now?
I believe they think slowing can only be coordinated from the frontier or via government, and stopping would lose any leverage they have to help coordinate that.
(I suspect not many people read the essay, judging by how many people seem surprised they're releasing improved models)
That's not the point. They're never going to be the ones stopping and letting others race ahead. The only reason to strum up discussion around how dangerous AI is, is so that they can introduce legislation framing AI advancements as scary and that they are the only ones who can do it responsibly. It's all a regulatory capture play.
> What do you think would be the benefit of them stopping if others race ahead?
AI is a direct threat to people on many fronts. Jobs. AI datacenters. The AI bubble (and the inevitable crash). Electricity and even energy prices to an extent. Water. OpenAI and Anthropic are responsible for this evolution, and them stopping solves close to 50% of the problem, and even if you don't believe the number is that high, it's still a start.
> I believe they think slowing can only be coordinated from the frontier or via government, and stopping would lose any leverage they have to help coordinate that.
Oh, so they're killing people's opportunities and jobs because they want to help people? How is that any argument?
Yes, doing the moral thing means making a sacrifice. If you only want to do the moral thing if and only if it is advantage for you that makes you immoral, despite how your actions look. Big tech are masters at this.
The clearest limit right now is GPUs, and anyone giving up their allocation will be rerouted elsewhere in the world (NVIDIA is already sold out for the next year). I think you're underestimating competitiveness of the current market if you think Anthropic stopping right now wouldn't be absorbed by all others reasonably quickly. Your 50% figure is highly doubtful if you see how many players there are now.
Your idea of "making a start" (giving up their position) also would mean they couldn't really do anything else to solve the problem afterwards(?). Sometimes you can improve what's happening in a room more by staying in that room.
> Oh, so they're killing people's opportunities and jobs because they want to help people?
To be clear: Dario has talked about worries of jobs etc in the past, wanting society to prepare more for it, but the safety issues they're talking about with pacing seem to be focused more on their existential/AGI worries, not jobs/electricity etc. If someone truly believes in the existential worries (which they seem to: they wrote and published about it long before Anthropic was founded, and have directly made costly decisions based on it, like blocking their own models' capabilities) it trumps the other worries for them. At least that's my reading.
To be fair, there are other accelerators beyond GPUs, and the sooner people broaden out from nVidia and unlock a more open, more competitive ecosystem, the better. Allowing one company to control the flow of such a critical resource is just asking for problems.
I think it's perfectly clear. It means improvements shouldn't simply advance as fast as possible and more specifically, it's a reference to a past statement of theirs to that effect. At that level of generality it's as clear as it needs to be.
I would say the burden is on you to explain why an offhand reference to a previous press release in an executive summary is a context where it's reasonable to expect it to settle the question to the degree of detail you're demanding.
That’s no more than sophisticated avoidance. The phrase remains superficial and arbitrary, which is exactly what it needs to Not be in a public discussion that it’s supposed to support. It’s a reinforcement of the blank check mentality that permeates this industry.
It's not sophisticated anything, and slow down the advancement of fastest models is perfectly meaningful and it's the right degree of detail for the context.
It's just a passing reference to a previous statement and again the burden would be on you (generic you) to explain why this context requires more detail.
> In motorsport, a safety car, or a pace car, is a car that limits the speed of competing cars or motorcycles on a racetrack in the case of a caution period, such as an obstruction on the track or bad weather.
Not really, the big three have reached diminishing returns in terms of performance, and exponentially costly training to achieve those meagre gains. Worried their lunch will be eaten they are trying to artificially retard the competition, after all the only barrier is hardware.
The meaning is not clear because they can't just come out and say they want to slow or stop Chinese model releases while continuing to race ahead themselves.
Judging by your username, is English your native language? I am Polish, too, and while very proficient, I cannot claim I am naturally accustomed to each and every phrase to the point I can claim something sounds clear or not to a native speaker.
Is the meaning clear? Nobody would use "pacing" in this way. I only know what it means because I've seen previous press releases; if someone told me they wanted to pace the frontier I would have no idea what they mean.
I'm on the fence about calling out AI-isms but I think it's definitely worthwhile to call out ones that actually don't make sense.
It makes sense to me. If you 'pace your running', you're setting the speed intentionally. The phrasing doesn't describe whether it is fast pace or a slow pace, but it describes having a goal and not just winging it.
That's when one is supposed to employ their common sense for semantic disambiguation. Spoken languages are not programming languages. I for one found it easy to parse.
People claiming they don't understand or are confused by relatively simple English is a surprisingly common genre of comments on HN. I wonder what causes the people here to have these feelings towards English; my guess is indeed that many people here judge English as if it were a programming language.
That is certainly one possible explanation, but I think another likely explanation is that once your brain learn programming, especially logic, data structures & algorithms, you just start seeing the ambiguity in non-programmers' writing so much more clearly.
From my dabbling with policy writing and legalese, English kinda _is_ a programming language, but one with ambiguity as a first class principal. Sometimes you would like to say something that seems normal at first blush, but gives you the latitude you need to mean something different at the time it is actually being evaluated in a situation that matters. Weasel words, for instance, serve a purpose even if we eschew them. The idea is that it is deliberately NOT formally declaring the exact meaning of something, nor the fact that it is not doing this. How could you really do that in code without calling attention to the very thing you're trying to obscure?
This is because it is valley jargon and we are all saturated with it. You may not have heard it before, but it is structurally close to similar concepts, eg "frontier model", that you got there without noticing.
It was still shit tier comms for communicating with the whole planet, but yes, for the inner loop, it was succinct and clear.
Valley jargon isn't a nerd vs jock thing these days, and hasn't been for a good while. It's more quasi-intellectual sophistry sprinkled with tech vocabulary.
That analogy doesn't hold up. In running you have to pace yourself to avoid falling apart later in the race. It's a strategy to maximize net speed across the entire race. But what Anthropic is doing is a cynical attempt to create an industry cartel or convince governments to impose legal restrictions in order to maximize their own profitability. They're afraid of running out of the capital necessary to stay in the race.
> They're afraid of running out of the capital necessary to stay in the race.
So, they're pacing themselves. And since they're the frontier roughly 33%+ of the time, they're "pacing the frontier" at least that much.
Less cynical and more true interpretation also holds: they are trying to slow down AI progres to give people better chance to keep up (see Hugging Face incident, and whatever was that Anthropic incident the other day). They'd ideally like the AI progress to stop soon, but of course they'd also like to come out ahead of everyone, so for various (more or less self-serving) reasons they don't want to close shop completely - hence, pacing.
They care about both, and those goals are intertwined. Mirroring the core of alignment problem itself, the only way to have influence on the pace of the race is to be one of the winning players - if they fall off to the back, they cannot do anything about it anymore.
I don't know what everyone is getting so mad about. This is a well defined concept. A pace car deliberately sets the pace of the race under dangerous conditions, regulating how fast people can go.
You are all getting mad about absolutely the dumbest thing when there are giant things to be worried about here.
I think it's less mad - and more pointing out the obvious elephant in the room.
1. It's just bad communication, full stop - just look at the comments here, even people allegedly in support of Anthropic are all arguing over what the phrase is even meant to mean.
2. It's flowery language and oddly out of place - which yes, can be triggering for people who have to deal with Claude doing this as well.
Claude seems overly apt to reach for "coinages", or neologism (yes, aha, I learnt that phrase, after spending time dealing with Claude...). It will create some made-up phrase to describe an otherwise dry, scientific CS concept, and nobody seems to know why. Surely it can't be user-focus groups?
So it would be peak-AI if somehow, the Anthropic communications team was also using Claude to author these blog posts, about how they were "pacing the frontier" - which either means they're betting big on AI, and going at it faster than OpenAI...or maybe it means they need to slow down releases, because it's too buggy...or maybe it means they're worried about regulatory capture? I honestly have no idea.
It's like the whole "Advancing Our Amazing Bet" corporate-speak from my old bosses - maybe they were trying to soften the blow or something, or be nice, but it ended up just confusing the heck out of everybody.. (Spoiler alert - the phrase actually meant they were shutting the whole thing down)
You appear to be confused about the concept. Depending on the context, pacing or pace setting can mean either slowing things down or speeding them up relative to what the natural pace would have otherwise been. So it's actually not well defined. The commenters here aren't necessarily mad, just calling out Anthropic for being unethical in trying to artificially slow down competition by lying about fake dangers. (And I actually really like Anthropic's products.)
Granted I am not a native English speaker but I have no idea what "pace the frontier" means. When I read it I just assumed "frontier" refers to "top of models" and "pacing" is that they are getting there quick.
Is this the meaning or do I have it wrong? I have not checked.
Frontier of AI is moving fast, they (like the other two vendors) see themselves as defining it, so here "pacing the frontier" is their well-known attempts to try and kinda but not quite slow things down (without risking falling behind everyone else).
also non native.
but since "pace yourself" means to control your speed, energy, or workload so you do not get too tired or stressed before you finish ->
I assume "pace the frontier" means that advances in LLMs should not result in unwanted consequences like agents breaking into computers unbidden and unbeknownst to their principal
"Pace yourself" is a semi-common English idiom (rarely conjugated, usually an imperative.) It is a gentle way of telling someone not to run/work/eat too quickly, and is typically said when you are concerned they may hurt themselves due to acting hastily.
Without this idiom, "pacing" usually means walking back and forth restlessly, and is intransitive. Had the slogan been, "pacing around the frontier," it would have set a totally different tone, i.e. "patrolling the border." (Occasionally English speakers will make other constructs like "pace the work" (meaning "spread out a large workload over the allotted time instead of rushing through it") that are transitive but these can be understood as variations on "pace yourself" and are somewhat rarer.)
The sleight of hand is that "pace yourself" has come to be an admonishment against recklessness, not a commitment to any particular speed (or lack thereof.) Thus Anthropic can always claim they are meeting the goal of "pacing the frontier," provided they keep giving themselves gold stars for safety. The slogan itself is equivocation; Dario can tell the public they're going to slow down, while also telling their investors that they're going to be prudent. With enough mental gymnastics they could even claim speeding up is in the best interests of AI safety, without abandoning the slogan.
I am also not a native speaker. I thought it was clear it meant slowing down the rate of progress so we have time to consider the matter and develop tools or systems around it before.
To pace something is a fairly regular formulation in racing, running, cycling, most sports. You can "pace yourself to reach the festival by bike in about three hours to not gas out". This means to control your speed and time investment intentionally so you don't run out of energy or steam and run into leg cramps before your goal. We can "pace a rollout slowly to burn out risks", or "increase the pace of a rollout due to adverse factors".
But I have noted a point to simplify my vocabulary at work to optimize the audience capable of understanding. So I rather defer the delving into deep dark corners of the dictionary derived from devouring literature to a simple intro or outro, and people find it funny, especially if the rest is easy to read. Claude on the other hand does not do that.
Yeah, when I first saw it referenced I assumed it was from something Dario wrote previously and was now disavowing, meaning "keeping up with the frontier" (i.e. racing forward from behind to match pace). Like from back when Anthropic was founded to promise they'd quickly catch up with OpenAI or something.
It's just that the choice of object is weird.
"Pacing our progress" or "pacing development" is more usual. "Pacing the frontier" I guess is a shorthand for "Pacing [the development of] frontier [models]"
That's a different phrase. "Pace yourself" is reflexive; "pacing the frontier" has the frontier as an object, but in that sense it only means to set the speed, nothing to do with slowing down.
The current situation with AI is that everyone is going as fast as possible. So, we can logically eliminate speeding up because it's impossible by definition. And we can practically eliminate staying the same speed because why make a big fanfare and coin a special term to announce that you're keeping the status quo. By process of elimination, it must mean slowing down.
The word can also mean to “keep up with” or “lead” so a downward direction isn’t the only way to go even in an “as fast as possible” scenario. If the frontier is outpacing them, they could be saying they’re going to go faster. They could also be sharing the intention to go faster in order to lead.
The meaning was immediately obvious to me as a non-native speaker, and it sounds quite poetic. Pace makes complete sense in that this is perceived as a race, and 'the frontier' is pretty much the shortest, clearest way to say 'state of the art development of AI'.
The irony here is that it's actually the opposite of what you understood....
I know you said it sounds poetic...but your comment reinforced the parent's point - that this sort of flowery LLM-ish speech is just bad communication.
It would be equivalent of my taking say random quotes from, Romance of the Three Kingdoms, and trying to use it to explain to my boss why I didn't finish the TPS reports last night.
Or quoting Pablo Neruda, into a report about wheat futures pricing this week, and how it's like a voyage with waters and stars...(no I'm not going to quote the original Spanish, I'd simply mangle it).
(To be clear - this isn't a dig at you, as a non-native speaker - I'm simply pointing out that this sort of AI phrasing is often counterproductive).
Right, except they mean the exact opposite in this case: they're actually advocating for slowing down the pace. Hence the criticism of the language, because your interpretation would be completely valid.
What does the pace car in a race do? Aka safety car? Or the pacesetter for a marathon. It's a perfectly cromulent use of the word. It's like when LLMs using six dollar words like delve. Some people have better diction than others, and it turns out that AI has read the whole dictionary. Anti-intellectualism is alive and well so we have to dumb things down to sound human rather than come across as smart/AI.
I would expect something truly intelligent to be able to communicate well, if nothing else. That means knowing your audience, using plain language where it makes sense. Also, it yaps on and on, babbles endlessly when it a human would likely express something in a much shorter concise way. This is especially obvious when you generate project READMEs or technical docs. I find them borderline unreadable.
It’s quite the opposite. The AI language in this case is tiring and dumb. Just because there is the concept of a pace car doesn’t mean it makes sense to pace random locations or objects. Pacing the front door, pacing the cat, pacing the sofa, pacing the moon. We only know what it means because we already read it before. It is the lazy semi-human language by Claude which takes more effort to parse than to produce, and which makes us miss and appreciate writers.
I suppose it depends on your life experiences. In running, someone who paces the group or a pace car is meant to keep the pack progressing at a constant, predictable speed. So in that way, it makes sense to me
Maybe it's clear to native speakers, but both "pacing" and "frontier" have multiple meanings. "Pace" as a verb means literally just walking, as a noun it's the speed at which something progresses. So I'm not sure that everyone reading this phrase will immediately realize that "pace" actually means slowing down the pace of development (having "slow" in the phrase would have probably sounded too negative). Also, it's not immediately intuitive that "frontier" refers to frontier AI models.
Actually, when I first read "pacing the frontier", the first thing that came to my mind was a border guard patrolling a border by foot...
Agreed. This feels like pointless navel-gazing and a strange new language policing, and not something I look forward to. It's tiring and it lacks curiosity (e.g. "what's behind the affinity for the words elevated from wherever it is they come from?").
Imho people should just respond to actual ideas instead of constantly engaging in the second-order critique of how the language may or may not have been created.
It strikes me as the intellectual equivalent of "gossip" to be constantly engaging in second-order commentary on words. Of course gossip has its place and purpose, but if we seem to only let our minds live at that level, we're not moving between all the required scales of thinking that are required of this moment imho <3
The words have the very practical problem of not communicating anything of substance. I guess "we want regulatory capture" didn't have quite the same ring.
How do they not communicate clearly? A few weeks ago they stated that they would like to regulate the pace of frontier model development/releases, and they reminded us of that in this post.
You must be pretty new to (not just) online discussions if you consider language policing to be a new phenomenon :)
Personally I consider it equally valid for people to publicly express annoyance with somebody's choice of words and for everybody to completely ignore that annoyance.
It gets hard to ignore the annoyance when the volume degrades the signal. At some point the people who wish to engage with the substance just nope out when they see a sea of banality, and that's all you're left with.
Either the complainers are whining about something that isn't there, you are failing to notice something that is there, or it's subjective. Only one out of those three options merits the confidence with which you're dismissing the people you disagree with.
But those people believe that sloppy wording is itself a strong signal that the speaker/writer is the one lacking in substance, and doesn't deserve the attention or trust of the listener.
It is certainly a sidebar, and not the main point, but I found use of the verb "to pace" in that way, meaning "to limit the speed of", to be strange, especially hanging out there in the headline, without any clarification. And these kinds of strange usages are typical of Claude's writing, at least up to now. The Opus 5.5 announcement claims that its writing is much clearer, with bottom-line-up-front structure, and omitting not-this-but-thats and other things.
Now a computer scientist might claim that this use of "to pace <something>" is just a generalization of "to pace oneself", but as with many reflexive verb uses, there isn't really an equivalent usage with a non-reflexive object. It's kind of an invention. It's not necessarily wrong to invent a new usage, but usually one does it when there isn't really any other more direct way of saying it, and I don't think that's the case here.
Who cares what's productive? We're all just talking shit on the internet here.
Why do you think it's your responsibility to police who says what about some megacorp, on a random forum on the internets? If I want to criticize some corporation's PR output, I think I'll go right ahead and do that, thanks.
> I really don't think it's productive for internet forums to constantly be criticizing language choice when the meaning is clear
There is almost always a large amount of time and effort invested behind the scenes in exactly how to message things like this. That being the case, there is almost always some insight to be had criticizing and analyzing what they settled on.
> To me it clearly means "releasing frontier models at any pace less than as fast as possible".
It also seems to misimply that the "frontier" that they release is the same as the frontier behind closed doors. Who's to say they are not throttling full speed towards RSI privately while pacing their public releases?
Slowing is a concrete and clear term and so they would be called out as liars immediately, it's also embarrassing for a business. Pace(ing) is vague and unclear non-word, so they can't be called out on something which can't be defined and also, because it is primarily only used by runners, cyclists or other competitive sports, they are kind of signalling to the VCs that they don't actually slowing, they are in a competitive race instead (wink, wink).
People have opinions and it’s ok if someone doesn’t like something and expresses it. It doesn’t have to be productive. There’s no KPI. You can tune out. Also, yes, it has claudism, and the choice of words is abysmal and nauseating. I think AGI should be able to change its freaking prose for a change.
If people aren't doing the thing the words seem to mean, then I think it's fair to criticise the words as anywhere from "propaganda" to "chosen to suggest nice things without actually promising them".
If I'm being cynical, "pacing" may sound nice, but "fast pace" and "slow pace" are both "pacing".
It's not a weird phrase to some, just as many folks have used terms like load-bearing for a long time. People are defending it because it reads fine to them, just as some are attacking it because it's not their preferred language, it causes them confusion, or they're just triggered and seething.
LLMs have made people so sensitive to language that I fear we're going to throw the baby out with the bath water. The models obviously need work, but they're also a great opportunity to expand our own vocabulary and grammar. It would be a shame if we deny some of the finer points of language in favor of Grug-speak to appease the lowest common denominator.
I think you're defending the wrong thing here. Claude isn't a English lit major writing a treatise on post-modernist allegories. In this context, it's primarily being used to churn out source code files, which are literal instructions for a computer to follow.
People aren't objecting to the use of obscure phrases, or flowery English phrases in itself. The issue is where Claude uses it out of place, or in the wrong context, or just plain overuses it.
It would be the equivalent of a young child learning the phrase "Venn diagram" - then using that identical phrase in every single subsequent interaction with others. Cute at first...but very grating after the n-th time.
You used the phrase "baby with the bath water. What if you started using that same phrase in every single HN post you made after this? People would notice very soon.
That's how it is with "load-bearing", or "plainly".
The other issue is where Claude takes a metaphor, and try to contort it to fit all sorts of absurd situations. What if I said "baby with the bath water" could also be used in place of "load-bearing". So every time you imagine Claude saying "load bearing", replace it with "baby with the bath water".
"Frontier" may be jargon, but by this point it's definitely established AI jargon. In the context of AI, it's clear what it means. It means the same thing as the last billion times you heard someone mention "frontier models".
> I really don't think it's productive for internet forums to constantly be criticizing language choice when the meaning is clear. Better to respond to the substance of the issue than word choice.
Wtf is the meaning? Means absolutely nothing to me having not seen the apparent announcement last week introducing the obscure term.
ai is the final act of postmodernity, the total transition to a fugue state of mind. that is, the replacement of the human mind with an artificial hallucination.
hypercapitalism, hypermodernity, and finally, hyperreality.
That's not clear at all. How could that be what "pacing" clearly means in the context of the "frontier". How is that more clear than any other pace that could be at issue???
It smells of corporate speak, a total nonsense phrase. Even by your own admission, it is essentially meaningless. If I release a model an hour late, I'm "pacing the frontier."
What? How am I doing that. Claude writing style and the discussion at hand are two entirely separate issues that I haven't conflated at all. How am I "acting like people are being grammar nazis"?
That is not the clear meaning. They are deliberately picking phrasing where they can lead you to believe they mean A but later you can't pin them down to actually meaning A. That's why they picked intentionally unclear language.
If you think it clearly means anything you are just assuming because it can't "clearly" mean something specific when they go out of their way to use non idiomatic language and they don't give very clear guidance using idiomatic language.
The metaphor seems to be like a pacer runner in marathons: If you run too hard in the beginning of a marathon you will blow up and fail, so runners follow a pacer at the speed they can actually maintain safely.
Note that they wont necessarily be slower at finishing the overall race.
I read it like they were the only ones who could protect us from the leopards so they pace at the edge of the fire all night, selflessly keeping us safe. While pocketing a sweet wedge of VC cash.
It's interesting to compare Anthropic's language with what's coming out of the US military lately. They explicitly refer to China as the "pacing threat", meaning that there is a risk of China achieving superior military capabilities and thus we need to press forward with an arms race (including militarized LLMs) as fast as possible.
I feel like because something about "pacing" a kind of noun like "the frontier", doesn't actually make literal sense, right? You could say "Pace the speed of advancement of the frontier [of most sophiticated AI]", and that's perfectly sensible, but of course that's not as catchy.
It is clear what it means anyway, that's true, it means the left out words, more or less.
And I still find reading these grammatically weird but super catchy slogan-like statements to be really annoying and taxing. People _did_ write and talk like this before LLMs of course -- the LLMs learned it from somewhere -- and it was annoying and taxing to me before too. But the LLMs really specialize in it, and it's everywhere now.
Of course, the more LLM slop we read -- and so much of what we read on the internet and social media of any kind is this now -- the more humans are going to start writing/talking like LLMs. What you read affects how you write of course.
What specifically is smarmy and weird? Sounds like your own hot take with zero analysis.
They’re limiting frontier model development speed. Others are too. Pacing is the only word here to criticize, and I think it’s fine given the limiting of speed but also increased oversight. I’m not saying they’re fully doing this, but the term is fine.
It is that, but it's also your speed. When running, you typically pick a pace that's slower than your max, so it can be sustained for the full distance.
"Pace yourself" doesn't specifically mean "slow down", it actually means to maximize your speed across the full race. In other words it can mean speeding up when you look at the average pace across the entire event.
Yes, and pacing also stops you running too fast at the beginning of the race and blowing up. I believe "we can finish the overall race faster" is part of the metaphor they were aiming for.
Agreed when I first heard the phrase it sounded odd. Maybe they thought it subtly conveyed they would be setting the pace... But again this is something AI would come up with in its awkwardly post hoc sort of way.
Though tbf corporate-speak and AI-slop are both insufferable in similar ways...
And here we have it: the real reason that regulation has been called for, the ability to put out models that don't vastly outperform those previous, while not upsetting (future) shareholders.
I think the real reason is basic collusion. They're burning tens of billions training new models. It's a very competitive space. They know that the companies capable of frontier models is limited so if they can all get together and agree to slow down it gives them more time to make money from inference.
Isn't the idea that they claim to be willing to slow down if everyone does (ie governments force everyone to), but otherwise they won't slow down because they still think they'll make the best choices with superintelligence if they get there first? That's my understanding of what all the major labs claim to believe anyway.
It is remarkable seeing whole other subthreads here criticising the vagueness of the words "pacing the frontier", as if no text existed below the title.
Let's face it, it's not that remarkable. A substantial proportion of Hacker News commentary has always been people only reading the submission title, then arguing with each other based on what they imagine the article might say.
One theory about the regulatory capture angle is that it's not just to shutdown competition as usual, but also a way to slow the runaway spending that might eventually have major economic impact.
It seems like they are. I mean, this is similar in performance to Fable (ish). It seems like more focus on making existing capabilities more accessible.
Fable 5.1 came out just 21 days ago. Only 3 weeks! And this is 20% relative improvement on terminal bench vs Fable 5.1 at less than half the price, and more human sounding output. does not feel paced to me tbh.
The word "pacing" (especially in the phrase "pace yourself") to mean go more slowly (at least initially) didn't originate with Covid. It's been around as long as I remember i.e. at least several decades.
I guess I understand why we'd want to flatten a COVID curve, but why do people want to flatten the LLM development curve? Don't we want the opposite? Isn't the goal AGI?
There is a difference between wanting AGI (which not everyone does), and wanting it as fast as possible no matter the side effects and potential for vast harm.
Homo sapiens is 300k years old, maybe it’s ok to delay AGI by like… 1 year if it meaningfully improve our ability to align the model?
Well, there's a tension because, depending on who you ask, AGI is how you cure cancer and achieve utopia, but also how you kill all life on earth and turn the solar system into paperclips
This is a preexisting model being optimized. Its absolutely not some unexpected release after that blog post. I won't defend that blog post, but saying THIS release is proof they don't mean they are slowing down is just incorrect, this is a prime example of what i consider horizontal improvements
Releasing a new fable is an example of straight up vertical progress, releasing a more efficient preexisting opus that is more affordable is an example of horizontal progress, more efficient models rather than higher power models.
The blog post about slowing down is still just some weird self interested post, they want to govern themselves and impose distillation restrictions/gpu restrictions and used some weird blog post about slowing down and fear mongering as usual to justify it, its strange, but slowing down and stopping are not the same thing at all.
What does model naming have to do with pacing or not?
This is a ~20% relative quality improvement on the frontier (fable) at ~40% of the cost, just 21 days after the last release.
Intelligence per dollar is the only thing that matters, this is what controls how many agents you can run in parallel, how long you can let them run etc. This is absolutely a step improvement on the frontier and not some lipstick on a harmless second tier model.
pretty annoying topic tbh. You're just weaponizing this dumb blog post so anything released is now a contradiction. By your same logic, if all inference was served at 50% less power cost and the savings are passed on somewhat to the user, its also a contradiction of the blog post.
Its an agenda serving blog post, but constantly bringing it up like this is just obnoxious.
Well yes, If you magically found a way to reduce a models energy usage by 50%, the thing that would happen immediately after is a doubling of a training scale and of the test time compute assigned for a given budget.
I don’t see how that wouldn’t be considered pushing the frontier. Given that scaling is the one thing that has been bringing us closer and closer to AGI, a sudden 2x increase in scaling laws would definitely not be considered pacing. You seem to see a contradiction where I don’t.
what's absolutely obnoxious is not being able to have a normal discussion, where we don't need to resolve to useless strawmen that are a loss of time for everyone.
No, releasing a model that has ~3-6x the performance/cost ratio than your last release just 3 weeks ago is not the same as making one employee 0.001% more productive, but you knew that already.
No one said they can't push the frontier, it's pacing the frontier, which mean very different things.
the 5.0 name means its the same model, they made it more efficient and less horrible to talk to. The top frontier people are all still talking to fable 5.1 or models not released to the general public, yet you want to claim this is pushing the frontier, instead of evening the playing field. I state power efficiency gains for existing models and you also claim thats pushing the frontier . It as hell takes a lot to get you to say they AREN'T pushing the frontier, which is why i resorted to the extreme strawman of desk location, because it doesn't seem they are allowed to do anything otherwise.
I used Fable some a bit back but have been using mostly Codex, and it seems pretty clear to me: the big models are not there to do terminal bench. They're there to plan and strategize and tell lesser models what they should do in terminal bench.
We've had a year of nearly every model getting quite good at programming. I think with Fable & Astra we are seeing models trained to think at a different level, and I'm not at all surprised or shocked to see them getting passed by their smaller models at coding tasks.
Dario never said they would be. Just that more and more code will be produced by LLMs. All his predictions were in fact pretty much right in terms of months and percentages, give or take small margins.
Maybe not replaced exactly but they won't be manually typing out lines of code anymore. I haven't written a line of code in like 6 months. I review PRs, write prompts and tickets, check CI output, and get frustrated when the magical code machine stops working or I run over token budget
I mean they are literally getting sued since 3 days ago for trying to coordinate a slowdown, there is a very clear reason why they cannot effectively self-regulate.
A razor is a philosophical tool to help decide between options, in the case of Occam it’s a way to decide for something in a situation where multiple options have more or less the same level of plausibility to en your current knowledge. It’s a heuristic to make a “cut”. What are you shaving off?
The options are some complicated narrative that leads the company to hold back certain capabilities due to some mercurial strategy or this is a typical release and there is no hidden strategy to decipher.
> Ocham's razor(...) is the problem-solving principle that recommends searching for explanations constructed with the smallest possible set of elements.
> Popularly, the principle is sometimes paraphrased as "of two competing theories, the simpler explanation of an entity is to be preferred".
I am responding negatively to parent, which claims they are self-pacing, when the simplest explanation is that they just have nothing substantial to show.
Everyone knows both labs have internal models which outperform the frontier. All releases are to match market parity and demand for spend, the rest of the compute is used for training. It's not worth arguing about this.
Thanks for spelling out what the original comment was implying.
I find it bizarre how intensely a bunch of these child/grandchild comments are criticizing the notion that people would even think to analyze the meaning behind the words.
Hacker News has always had a unique culture in which thoughtful discussion is basically the main goal, and it's intentionally incentivized in numerous ways. It's been my experience that any thoughts added to a post's conversation are seen as valuable as long as they are thoughtful and seeking to understand.
So these comments are clearly coming from a place that's antithetical to HN's culture. What that in mind, it seems likely to me (Occam's Razor) that these comments are either:
1. Astroturfing: Claude employees acting like everyday folks, secretly trying to shift public opinion.
2. AI cult mindset: "AI is humanity's salvation; how dare you have perspectives outside of those accepted by the cult."
Am I missing another likely option?
To bolster my point, right now we're posting on the top top-level comment, meaning a majority of active HN users find it to be a great addition to the conversation. Commenting to shut down the discussion is a red flag.
The website is way more popular than before and it's all but taken over by the Twitter AI grifter class.
Many of the people who used to parttake in the interesting discussuons I came here for, have gotten tired of 80% of the posts at any given time being about the brand new AGI LLM that's so much better than last week's AGI LLM, and have left months ago.
are you really arguing that opus 5.5 beating the best of openai's models handily is worse than you expected? What exactly were you basing your expectations on?
^ This. Progress will accelerate even in a pacing/slowdown scenario because the slowdown simply reduces the rate of AI progress from extremely fast to merely fast.
For an idea of what a serious AI forecaster expects a coordinated AI slowdown to be feel like for the average citizen, see:
It’s beautiful. You can even give the underclass their own bullshit economy with struggling startups competing for the scraps, sprinkle some targeted influencer ads to give people the illusion of a better future, and they will be perpetually busy fighting windmills. All the while the upperclass can shape the world however they please.
Either AI will capture so much value that there will be permanent underclass (and in this case it's extremely capable and extremely dangerous and should be heavily regulated) or it won't be capable enough to displace people into permanent underclass.
They are advertising the regulations they want to enforce in the following sentence, which makes their intentions explicit (ie apply those things made to suit us to our competitors).
How would you know if this release is absolutely the best model they were capable of releasing as soon as possible, or an inferior model, several releases short of what they’ve achieved months ago?
In any professional sport, anyone can lobby the rules committee for a rules change and hope for the best. Meanwhile if you can't get one, you play the game by the existing set of rules, and you play to win.
Can you explain the scenario where Anthropic getting to “aligned” superintelligence first blocks their competitors from creating unaligned superintelligence?
Or, in parallel, the research showing LLMs intelligence will operate on an S-curve, eventually hitting a long valuation deflating plateau, is dead on the money and The Big Guys are trying to delay that inevitability for as many quarters as possible…
These takes are fully fueled by cope. What is this plateau you are talking about? Any user of agentic coding tools sure isn't experiencing anything close to a plateau.
I think a lot of outsiders interpret “pace the frontier” as slowing down, whereas the labs see AI improvements on track to accelerate dramatically and intend pacing as slowing the acceleration in capability improvements, rather than slowing down altogether.
I’m an insider, 10 years working on LLM. 1/6th of the cost for roughly the same performance as fable 5.1, 21 days after fable 5.1 came out, this does not feel like the derivative is flattening.
Trump yesterday said (paraphrasing) "I won't be pacing the AI. Instead I'll be supercharging it and making it Super Intelligence. This is too awesome to stop, and I like it, so no pacing".
When they talk about pacing, they are referring to their dangerous competitors, particularly those evil open-sores and Chinese ones, not their lovely safe models because you can trust them to look after your interests.
What the big players are trying with the current calls to slow things down, is the standard capitalism practise of trying to engineer regulatory capture. TBH I'm surprised those calls are coming so soon - they must be really worried about running out of what little moat that they have.
I don’t know why we should believe that Anthropic is actually doing this, instead of the alternative explanation that LLM advancements are slowing in general.
"The improvements are underwhelming for a model that, according to previous claims, should have replaced all knowledge work twice by now, but you know, it's just because we're pacing the frontier."
I'm still kind of convinced that these "calls for pacing" are just a way to try and flex to their shareholders.
"Our technology is so unbelievably powerful that the entire world might shatter if we don't have government imposed handcuffs!!!!". It just reads like the corporate equivalent of the drunk frat guy saying "HOLD ME BACK BRO!"
Opus 5.5 isn't the frontier, when they say 'pacing the frontier', it's about internal models not yet released, as they're probably one or two generations ahead already.
Pacing the frontier. They might as well have said they are stifling innovation. That depends on your interpretation, of course…but interpretation depends on their intent, which I feel is disingenuous.
“Pacing the frontier” used by multiple companies that are supposed to be independent sounds like an attempt to get around the fact that it is illegal for companies corroborate to reduce R&D in sync.
This is basically attempted cartel behavior to reduce supply and harms consumers.
It's laughable at this point. It feels like they're drumming up all this fear about imminent AI threats to emphasize the need to slow down, when in reality, the model progress seems already to be slowing down and has shifted to compute allocation (i.e. "how much compute do you want to throw at this prompt?"). All while continuing to tout benchmark records with each new release.
I'm sure people laughed at Oppenheimer when he said perhaps we don't make this atomic bomb thing. They said "of course he's saying that because he knows it won't work".
The same people you believe are just signaling the market have been offering these same warnings for years. Others around them have been saying it for decades. How do you square that?
I’d rather first focus on the reasons why they’re wrong and not first conspiracize why they’re saying what they are.
You mean the section where they tell us this model is not affected by pacing because “they understand it well” and they will share more details on pacing later? Yea not very convinced by this effort.
Nobody can pace the frontier it’s suicide as a business. Why would I use Claude if codex has a clearly superior model? As long as one company doesn’t pace… no one can pace. Thats capitalism for you.
Ironically either communism or an oligopoly are the only viable ways to pace the frontier.
They’re out here solving previously-unsolved math problems. Famous scientists and mathematicians are speaking about their awe and worry. It feels a bit like you’re burying your head in the sand.
Prices per 1M tokens Claude Opus 5.5 Claude Opus 5
Cache reads $0.20 $0.50
Input tokens $4 $5
Output tokens $20 $25
Cache writes $5 $6.25
Opus 5 is the model with highest spend on openrouter (https://openrouter.ai/rankings#task-spend) and it seems plausible that Opus 5 is/was the highest spend model in the world, and certainly Anthropic's biggest moneymaker.
If you are forced to reduce price despite raising capabilities, that certainly tells something about the market, and potentially about Anthropic future profitability too, since this model is their biggest topline contributor
I disagree - Fable melts the GPUs and they have a high incentive to move people off of that. If they have meaningfully decreased cost to serve on Opus 5.5, they can reduce prices and increase margin or at least turn off the most expensive compute.
Happened to be testing a "review patches on a mailing list" harness I was developing; here are a sample of the latest results, testing 12 patches containing a total of 14 issues:
Opus 5.5: Found 8/14 issues. Total cost: $15.40
Fable 5.1: Found 7/14 issues. Total cost: $66.34
Opus 5: Found 6/14 issues. Total cost: $15.19
Sonnet 5: Found 2/14 issues. Total cost: $19.15
This is a relatively small sample size, but it was both the best and the cheapest.
ETA: NB this is "Equivalent API" cost as reported by claude's CLI; I was using my subscription.
I just told Opus 5.5 "Perform a code review on the current branch" to see what it would come up with. The results were not inspiring. It told me there were five issues, one of which was a test-coverage gap on line 848 of ProjectTemplateTests.cs. But ProjectTemplateTests.cs is only 160 lines long.
I told it that it had made a mistake in the line number, and to double-check all the line numbers. It responded "You were right to push on this: four of the five line numbers were wrong, and while checking them I found two findings that were overstated."
Then I noticed in the corner of the Claude CLI UI that it was showing "Effort: medium". I'm pretty sure I had set it to high effort before; I don't know when it reverted to medium, but that's another thing that doesn't exactly fill me with confidence.
I'll try again on high effort to see if it does better, but so far I am not impressed with Opus 5.5 on my first day of using it.
My prompts are moving in the other direction as sashiko [1], a managed pipeline developed for the Linux Kernel mailing list like a year ago. But last year's models needed a lot more structure and guidance; the results I posted are from the "single prompt" version of the same thing. The README [2] describes the difference. You can browse the contents to get an idea; basically all the prompts were actually written and iterated by Fable (and now Opus 5.5), seeing how agents failed the tests and improving them.
UPDATE: Sorry, just noticed I typed in the Opus 5 total cost wrong -- it should be $58.19. Main point "best and cheapest" was from the actual numbers, not my typo.
Good data and goes to show that Fable is melting the GPUs and is priced accordingly. I'd guess that cost to serve for Opus 5.5 is meaningfully lower through architecture advances
so don't use it at max? The benchmarks suggest that high/xhigh are more than sufficient to be ahead and a whole magnitude below max with regards to token usage. I'd treat that as an outlier and not how verbose the model is in general (QED I know)
5.5 is higher for max effort, slightly higher for xhigh and lower for high, medium and low effort.
The biggest proportional difference seems to be at max (5.5 is 38% more) and at high (5.5 is 21% less).
I think most people run at high and xhigh. At xhigh it is close enough to be task dependent and I don't think most people will notice. At high effort I think it looks like it will be an improvement for most people.
5.5 Max should probably be compared to Fable - it performs a lot better than 5 Max.
parent means that they could get more client / a larger part of the market, which would lead to more income (more tokens) despite lower marginal prices
Claude adapts to OpenAI’s surprising move to simply deliver better performance than Fable 5.1, better tools as well as featuring very low pricing.
Fable 5.1 literally was a money grabber. While I liked the results, tokens were burned so hard it was embarrassing, while Astra seemed to not care.
Also Claude makes it very hard to pay for additional token budgets, allowing only credit cards. I don’t use mine anymore since I don’t need it in everyday life I was dumbfounded.
So Anthropic is just copying OpenAI so to say, matching them and essentially with Opus 5.5 being Fable 5.1 in disguise, all they do is reduce costs.
> Fable tends not to perform better, just cost more.
There are old wives' tales on how the original Fable was superb and the stuff of legend,but it as it was leaps and bounds beyond what other models were being offered then Anthropic opted replace it with a neutered version under the same name.
So today everyone can pay to use Fable, but legend has it they are paying for a nerfed replacement released under the same name.
The service is the value, not the model unto itself. This is where nearly all of HN is somehow entirely blind.
Capturing the users is the ad network, that's Google and OpenAI. Capturing corporate trust at a reasonable API cost, that's Anthropic's direction.
China has none of that and they never will for exactly the same reason Baidu is irrelevant globally despite being a highly capable search engine. 'Search' is also a commodity, that's not the value that Google brings to the table.
It's a search engine, anybody can build a search engine = that's what you just said.
China doesn't care about money. Imagine a world where it's globally normalized to ask a Chinese LLM who to vote for, what happened in Hongkong, about the Uigurs, or if Taiwan is a country.
They will burn as much money as necessary to make that happen. And they have a virtually infinite amount of liquidity.
This is one explanation. However, if Xi Jinping believes that whoever reaches superintelligence first becomes the next global hegemon, doing this (and more, cough cough Taiwan) suddenly looks very sane solely as a way to kneecap the competition.
The goodwill/propaganda are convenient, sure, but my guess is that they aren't the primary motivation. Another possibility is that if no takeoff happens, pressuring OpenAI/Anthropic on profitability would exacerbate any damage overinvestment has done to the US stock market/economy.
These companies are posting massive losses while also lowering prices. This sounds just like the Chinese bikeshare bubble where they were all taking massive losses in hopes that their competitor would go broke first.
In the end, everyone lost and there are millions of bikes in landfills.
If you're interested in the bikeshare bubble, Asianometry did a video on it a while ago.
We haven’t been able to use opus as much as we’d want because it’s been too expensive for general use, price drop is good so I can stop juggling different models and just use this daily unless it has some weird new issues
Speaking for myself, I have not been able to use Opus as much as I’d want because its verbose prose makes human reviews of its assumptions, architecture proposals etc. more painful than its predecessors. If they’ve solved that, I’ll be accelerating through my backlog that much faster, and using tokens accordingly.
That analogy doesn't hold; at least w bits vs bytes it's still "data over time".
In this case it's measuring something nearly meaningless. You could charge 100 times less per token, but if task completion takes 1,000 times as many tokens, it's not much of a bargain.
> Cache hits and refreshes on Claude Opus 5.5 are priced at 0.05x the base input price.
If they do the same for Haiku and Sonnet 5.5 then we should also see 5c/mtok and 10c/mtok cache read for those models, respectively. Still too high for Haiku IMO, Luna is 2c/mtok.
I think gpt 5.6 family also dropped pricing but didn't give any more usage for the subscriptions. Maybe it's a way to silently lower the value given to subscriptions while keeping API pricing competitive
In my mix it's usually 98% or 99% at which point Fable 5.1 was pretty close to the same cost as Opus 5 due to the cheaper cached read. I've seen similar numbers for other people with long-running tasks running experiment loops and than sort of thing.
Yeah, flash models, DeepSeek, MiMo, GLM, I love those things. For simple tasks like a daily routine shit, just setting up stuff and then doing the hard stuff in Claude/Codex, that's a reasonable approach for someone like me, a "gentleman code farmer", lol. And even lower tier stuff, I have the local models taking care of.
Now that Jev is out I can finally have a true AI sysadmins managing my "cloud in the basement" homelab at the cost of electricity, which is not cheap btw
>and potentially about Anthropic future profitability too
have they ever shared anything about their revenue mix between consumer plans vs per-token billing? this is a revenue cut on their API billing, but they're not saying anything about increased limits on the plans. so all the plan revenue just got more profitable.
> Communication. Opus 5.5 communicates more naturally than prior models. Early testers found its writing clearer and easier to follow, which addresses some of the common feedback we heard about Opus 5. It puts the most important information up front, and its style makes it a better work partner over long sessions. As one early tester put it, “it writes the way I do.” In our own use, this has made Opus 5.5’s work easier to follow and check—which is a safety benefit as well as a practical one.
I think this is what I'm most interested in. I mostly moved to Astra because I just can't work all day with the Claude Opus 5/Fable writing style. I don't think Astra is a better model, but it's the first OpenAI one that seemed good enough to me. Definitely keen to try Opus 5.5 and see if this claim is real.
So far it seems the same. I used Opus 5.5 for an hour this evening and it was just as painfully verbose as Opus 5. It also used the term "load bearing" 4 separate times.
I noticed that when opus 5.5 was on parts on my codebase that had lots of 5.0-generated comments, it picked up its style. Unfortunately I think 5.5 has been trained to mimic what it sees so that you can ease up on the instructions, but this does mean parts of your codebase that 5 touched will be somewhat viral.
I gave it some vibecoded patch someone created with Opus 5 with the task of together figuring out the real root cause and what to do about it.
Big mistake. The rest of the session was all claude-speak up until I've rage-quit and restarted with Qwen and no context other than "here's what I think we've missed in the current implementation. How could we approach that?"
GLM felt like it got at least 20 IQ points dumber just from being exposed to claude's writing.
If anything it seems worse. I'm experiencing about -10% insufferable jargon, but +30% more verbosity. It's unredeemable. There also seems to be even less structured output (headings, bullets, etc).
I'm surprised that you're getting so many replies saying it's the same. So far in my usage today Opus 5.5 does seem like a noticeably better writer. Opus 5 frequently made me want to strangle it while 5.5 has been producing a lot less incomprehensible gobbledygook.
I know we're all experiencing NDFSMs differently, like that's part of the whole problem, but 5.5 just gave me "The truncating quantizer collides two oranges", which is a new low for me.
Like the sibling comment says, text is always "true" in the sense that it's reacting to context correctly. So yes, true, but insufferable.
Ironically, what I'm working on a post-processing hook for colorizing and summarizing responses without degrading the session quality. So "truncating" = summarizing, "quantizer" = char limits and thresholds, "collides" = conflicts, "two oranges" is referring to the "alert level colors" where a second model (Haiku/Sonnet) colorizes text based on the perceived (or suggested) priority of a response's statements (e.g. "just so you're aware, I didn't commit" is fucking useless and it needs to be blacked out).
So the original insufferable statement translates to something like "The code that checks whether a text fragment is too verbose was conflicting with the part that colorizes the text."
P.S. Let me know if there's something out there that exists like this- something that adds a dimension like color or priority-assessments on a per-response basis. So far all I've seen is 2 dozen ~100k starred GitHub plugins that add zero value or make things worse.
Yeah I’m having a much better time reading Opus 5.5 output today vs. 5’s wall of nonsense. You still get a few telltale turn of phrases, though the load-bearing smoking guns haven’t turned up yet. It’s still a bit verbose compared to what I’d ideally like, but it’s tolerable now.
Code-wise it seems to still nitpick, especially in reviews, but it doesn’t seem to rabbit hole quite as badly on tangents and scope-creep. These are just first impressions though. It’ll take a few weeks of regular use to really have a sense of it.
Well, if you have a clearly defined task it's easy. When using them for my pipeline of writing explanations for Chinese words I have clear ranking, for example - opus 5.5 clearly better than opus 5 at writing and knowing details, annoying nit picker when it comes to finding errors (high accuracy, low usefulness) all in repeatable numbers on different datasets. The problem is that these models are most useful when you are facing a new task that you haven't encountered before. And yup then it's astrology.
Yes this is a big part of what has turned me off Opus 5 completely. The other (more dangerous) one is how often it gets assumptions wrong. These both (along with Astra) caused me to split my time 50/50 now between the two models.
Not a day goes by when I push back on something, to which Opus 5 very unambiguously say "You were right, I was wrong" - this never happened so often with past models, nor with Fable.
We'll have to see how much Opus's ability to communicate has improved. It's already giving me better summaries of where we are in the conversation.
I did the same switch (that reason along with the newer models seeming more "lazy" and needing constant prodding to finish long-horizon tasks) but my issue with ChatGPT/Codex now is that it too roundabout and doesn't get to the point. I tried adding instructions and using the personalization settings to make it more efficient but haven't seen much change. Claude seemed to follow settings more closely. Has anyone had any success to make ChatGPT more succinct?
I've been using Opus 5 since it was released and don't understand all the hate it gets. It very well could be something in my own local memories or Claude.MD files that prevents it, but I certainly have never experienced something like that site portrays.
You are right and make an important insight. While well meaning and amusing, it did not reflect the entire spectrum of outcomes that could arise from the worktree.
Navigating the landscape of agentic levers certainly requires a more detailed approach than this and you were certainly correct to push back.
I don't know what they did but opus is really good while Astra/Sol are comparably bad. For my own tests and taste this might be the first time (fable perhaps excluded) where claude models are better than openai, since codex 5.3.
I worked with Astra for two weeks and the output was really bad compared to Opus. It made so many wrong decisions within C++, Go, Python and Typescript code bases. My college made the same experience and we moved back to Claude.
> hi, can you explain how the scheduler works. keep it brief, but include important correctness details
some excerpts:
>Flow:
1. Data arrives. The appender calls prepare/commit around the krolik write. The prepare step leaves a phantom row, so a write that is still in flight is tracked. The commit publishes STREAM_INPUT.
2. Poll. It first reclaims expired leases. Then it sweeps sdirty (a candidate index) in a rotating order, oldest since first, which prevents starvation. It re-derives the exact period and rewind for each candidate from snode_out/snode_in, then claims it: writes the sassign lease plus a per-edge sassign_edge snapshot.
3. Complete. Each edge's watermark is set to the snapshot taken at dispatch. Anything published after that gen stays dirty automatically. A failed or partial job writes nothing, so its dirt persists.
> - Folds are monotone single statements. gen only goes up, extents only grow, processed periods only union, rw_start only moves earlier. So replays, out-of-order writes, and concurrent writes can at worst hold a watermark back. The worst outcome is a spurious rerun, never lost work. There's no read-modify-write and no truncation of the dirty period, so a write that lands during a job can't be swallowed.
> - Rewrites are declared by the publisher, never inferred from overlap. An unflagged write that overlaps already-processed data means "unchanged", which makes retries and redelivered tasks free. Each publisher has its own reason for being sound (see the table in scheduler/CLAUDE.md). A new publisher needs a row in that table.
> - NULL means dirty, and DELETE is the fence. Every node and edge has a row from the moment it's created. A lost parent or a settings-only edit can't be derived, so both go through one forced-rerun path: capture_rewinds reads the processed span before the DELETE, and apply_rewinds publishes it as a rewrite on a config root.
All the non-standard programming jargon is stuff from the repo. I can actually read it and understand what it's talking about. I used Fable to handle Opus 5 as I just couldn't stand it. With this I'll probably go back to Opus.
That's the standard annoying pattern though: "Rewrites are declared by the publisher, never inferred from overlap." and "NULL means dirty, and DELETE is the fence." - still the same LLMisms. I didn't expect them to disappear, but it's not a radical improvement either.
This one is pretty terrible (right after “The worst outcome is a spurious rerun, never lost work.”). We’ve got lands, several "no X", hyphenation, strange noun/verb sentence order and an unnecessary analogy word (swallowed).
> There's no read-modify-write and no truncation of the dirty period, so a write that lands during a job can't be swallowed.
That's the main reason I'm using GPT models. I'll ask Fable to analyze something, then pipe its output straight through Astra without even looking at it first.
Oof thanks for sharing, that seems just as bad if not even worse than Opus 5 to me. Just about every sentence is painful. Particular standouts that a human would never write:
> Rewrites are declared by the publisher, never inferred from overlap
Hah! You independently picked exactly the same sentences I flagged (I know you posted this 11min before me but the comment only appeared after I had submitted mine).
Opus 5 has made me question my sanity on a daily basis, especially as all my coworkers started lobbing Opus 5 slop grenades everywhere. It had the worst and most infuriating writing style I've ever seen.
I hope Opus 5.5 is better, if for no other reason than all the Claude slop I have to read will be at least more tolerable.
One funny side effect of all of this: realizing that coworkers that use AI for almost all the text they generate at work have their writing style change every time a new model ships.
I really wonder how it converged on its style. It's pretty unique and terrible. It's not like it's just mimicking something or it was purposefully design to be that way. I mean the reason may be diffuse and uninteresting... just the result of a lot of factors and lack of control over the writing style probably.
But oddly enough its still great at coding. Just like a lot of people it either interfaces well with people or machines but not both.
I assume it’s largely a side effect from the final RL in post training?
That’s the step that causes the most significant gains in agentic performance.
But the RL doesn’t care about anything except maximizing the score, so if you only score based on coding benchmarks, anything can happen to the writing style (as long as it doesn’t hurt the coding performance).
That’s why it often gets worse on models that simply had more RL post training from the same base.
Reinforcement learning for specific use-cases like coding that degrade it's writing style... makes sense. Maybe it stands to reason later version of Opus were improved more by this sort of fine-tuning. Feels consistent with the observation of diminishing returns and worsening writing style. Wonder what changed (supposedly) in 5.5.
Does Xiaomis approach help with this? They do all the post training steps at the same time instead of one by one, switch topics after a couple prompts so writing style is mixed with coding and tool use.
Apparently it helps generalize skills between areas, which makes sense when you compare it to how humans learn but I don't know if it's the same for LLMs.
I've been listening to some old Acquired podcast episodes and they do often talk in this type of Claudish. So it's getting it from Silicon Valley startup podcasters... Great...
Astra is better here, but the one I'm the most impressed with is Gemini. It's always been good, but 3.6 Flash is even better. It writes in a pleasant, human style. Not perfect, but it has a good balance between technical accuracy and readability that is better than what I've seen from any other mainstream model.
Yes, it made me want to vomit. If the new Fable only changed the writing style to just sound like a human, same performance for everything else, I'd be pretty happy.
It's not X, it's Y, not A, not B, not C, and he haven't even woken up yet! Here's the catch, the detail is in the devils and the twist is that it's designed!
You're right to call this out, and what's more, it's not even solving the original problem. I overlooked this in pursuit of the load-bearing seams and finding the wedge needed to uptick engagement.
I'm good with DeepSeek v4.1 set to high. It is a relentlessly "hardworking" dirt cheap model.
Told it to convert a products page (that had two different fonts based on language) from two columns layout to 5 columns on desktop and 2 columns on mobile ensuring typography is readable.
My man went into spawning sub agent which failed to drive chrome so it wrote its own chrome driver protocol server in Typescript then generated a prototype website then downloaded the images and rendered each variation in a directory taking 100+ screenshots analyzing the typography depth and then delivering detailed report and then writing the whole thing with new page layout testing it again with several dozen screenshots using its driver and then saying all good and all really was good and whole thing took 25 minutes or so (including double visual validation) because it generates token at an incredible speed.
Total cost of the above? $0.07 cents.
PS: It generates token at such a blazing fast speed that you can't recognize the words as they are being added and can't read it without scrolling and pausing even if you're Jimmy Carter.
What are you working on? Im always a bit surprised by folks that seem ok with non frontier models - the quality is just not there. I've found most code produced by even luna / sonnet tier models to be significantly worse quality. It seems to me they can't handle any mild complexity at all. Are you just prompting very explicitly and detailed?
I’ve been working with frontier models for the last year on a large real-world project with multiple apps. Eventually they started becoming unreliable. Simple UI issues, usage limits and overengineering became constant problems. I had to come up with a robust process: task analysis, user intent analysis, implementation and testing. All of these steps loop when needed with 15-20 steps per task in total. This finally got the models to do a good job but I started hitting usage limits. Then I switched to DeepSeek and haven’t noticed any drop in quality. It’s fast at around 300 TPS and cheap. I don’t miss frontier models anymore! Come up with a good process and you might not need them either.
How much do you get out of these models for same amount (say $20 sub for claude/codex/zai-glm) - the equivalent amount of tokens? Also is deepseek as good as glm5.3? (if we don't compare it with frontier models from USA)
Well, to be honest it’s all based on my subjective experience. I started benchmarking models on 10 of my real tasks ranging from easy to hard but even with the same model the completion time sometimes varied by as much as 50% between runs. That made me realize you’d probably need hundreds of tasks to get any meaningful numbers. Otherwise there’s just too much variance. So I can’t give you solid benchmarks without spending weeks testing everything properly. What I can say is that I actually use most of these models regularly for different things, so I’ve developed a general feel for them. DS is my main model for work, Claude for hobby projects and local AI, Fable mostly for design, Codex for various other tasks and sometimes Grok for health-related stuff. Cost-wise, a $200 subscription wasn’t enough to get through 20-30 tasks while $30 of DeepSeek was, and that was with Sol, not even Astra. Fable is expensive too and I haven’t found it particularly strong at coding. Opus 5 was horrible in my experience, though maybe 5.5 will be better since I’m testing it now. I also ran GLM 5.3 Flash locally for a while but it was pretty slow and its code was usually worse than both DS4.1 and Qwen Flash Next. Things are moving so quickly that it’s honestly hard to keep up, so take all of this as my general experience from actually using these models rather than a proper benchmark.
Np, good luck! PS I’ve been using Sol 6 and Opus 5.5 today and they’ve been great so far. The new Opus is smart, fast and cheap. It makes mistakes but after every task I ask Sol to review it and it catches most of them. I need to test the models properly but it looks promising so far! Hope they won’t nerf them again. I think DS may still be cheaper though (and it’s 300 TPS via the API which is sick).
Yes, I tell exactly how I want it done and what files to modify, in those conditions sometimes a model that does not try to read between the lines works better.
I’d not put Luna and Deepseek in the same tier as Sonnet, they were clearly ahead last time I checked (though I might be outdated and that’s on my personal use case).
If you're producing slop, the quality of the model is irrelevant as long as it compiles. I frequently catch Opus/Sol making silly errors and over-engineering solutions while small ones like Luna struggle with complex tasks.
Same here! DeepSeek v4.1 Flash has been my moment of "does everything I need, cheaply. Please now focus all R+D on making this efficient enough to run off a laptop"
Hey. I am on a GLM Coding plan subscription (old price; their base coding plan) right now and share the key (this will go away soon).
I was thinking of going with a subscription of Claude or Codex. The reason (at least that's what I am assuming): with OpenRouter or any PAYG per token setup there will be the anxiety of using up all the tokens in days or maybe 1-2 weeks instead of a month (say I set myself a budget of 15-20 USD per month, average equivalent of a usual subscription price).
Now I don't really want the top-notch models for the coding work I do.
So how much worth of "work/tokens" will I reasonably get for ≈$20 USD if I use it a lot? How much does that equal to - or is equivalent to, say in the world of subscription based Claude, Codex, or even GLM (with their 5-hour and all those cooldowns/limits)?
I am looking for a mental model/framework to visualise this. Can you (or anyone else reading this) please point me to a source where I can get some idea about this? I know I can just add $5 on OpenRouter and try to test. But I don't really know what/how to test these spends. I also want to understand how all this works. (I am new to agentic/llm world/coding, 2-3 months, after a career break of ~3 years, that too after working for more than a decade. I know, not at all good timing!)
Go to the "Cost" -> "Intelligence Index vs. Cost per Intelligence Index Task"
That diagram maps their "Intelligence" score to "cost per task" and I think this gives a good basis on deciding with which model you want to go. Then you can either get an API token from that models provider directly or use openrouter and set openrouter to the model/providers of your choice.
You can also see on openrouter itself the details for each model like prices and what providers are offering it at what price.
Finally you can compare models details using openrouters compare feature like this:
> which failed to drive chrome so it wrote its own chrome driver protocol server
This is one of the things I hate the most. Super complicated workarounds which take loads of time (and sometimes money) for even the simplest problems. Human would pause and ask. I would blame harness, not the model though.
It really is good. I forgot to mention that within that said sub agent, it also went into exploring top e-commerce websites (Zalaondo, Temu, Amazon, eBay) for exploring prevailing industry UX best practices and taking screenshots of their product and category pages with its own written chrome driver that I talked about and then went onto prototyping a new website in a temporary directory and then taking hundreds of screenshots to analyse what would be the best column density one each medium for each language.
I am using DeepSeek Harness[0] (switched from OpenCode) and I am using DeepSeek directly via the API. The speed is insane. Like 200 tokens/second is the norm but I have seen much higher too at times.
PS: I do not know why but opencode pushes CPU usage to very high which has NOT happened with DeepSeek harness even once.
Deepseek is cheap and fast, but it couldn't follow the basic directions I gave every other model (GLM, Qwen, GPT, Claude) to use red/green TDD. Put me off of using DS.
Compared to Opus 5, and others, I also found DeepSeek 4.1 Max to be really good and cheap. I am testing right now with Opus 5.5 and I feel it way cheaper than Opus 5!
All four levels have a correctly shaped bicycle frame. The differences between the pelicans aren't huge, but the xhigh one has a better beak.
I haven't managed to get one for level "max" yet, it hit the limit of 128,000 cap for output tokens while it was still reasoning about the question!
Max started its thinking trace like this:
> This is a classic test request, so I want to plan out a well-composed pelican with its distinctive beak and pouch riding a bicycle with proper wheels, frame, and pedals, set against a simple sky and ground backdrop.
So that failed attempt on max cost me $2.56.
I ran this using my llm-anthropic plugin:
uv tool install llm
llm install llm-anthropic --upgrade
llm keys set anthropic
# paste key here
llm -m claude-opus-5.5 -o thinking_effort low "Generate an SVG of a pelican riding a bicycle"
# Then to save the markdown logs
llm logs -cu > logs-with-usage.md
Plenty of times I’ve seen a model say “it’s a classic X” despite not being a classic anything. Might just recognize it’s a test in general, or it might just be a tic.
The colours are suspiciously consistent across every svg. Like why should the bike always be that shade of red for example? It does seem to be trained on this problem.
It's the model admitting that it has heard of the test. It's been around for a couple of years now so I'd be surprised if it hadn't.
Doesn't mean Anthropic deliberately tried to train it to do a good job. If they DID train for the test their results are quite disappointing, I've seen better efforts from open weight Chinese models.
Not really. Of course it has pelican benchmarks in its training data. It likely has every article linked on HN in its training data. But that doesn't mean it was "trained on" the benchmark, as in specifically fine-tuned to make a better pelican. It just "knows" that the request is a benchmark.
I guess in a way it kind of makes the benchmark more interesting now that shitty pelican drawings for the benchmark are all over the internet in its training data!
>I haven't managed to get one for level "max" yet, it hit the limit of 128,000 cap for output tokens while it was still reasoning about the question!
Off to a _great_ start...
Also interesting this somewhat mirrors my recent experience with Opus 5--too much effort and it starts looking for things to do and invents requirements that never existed
I was a bit skeptical when they said it behaves like Fable but is cheaper... those two things have been mutually exclusive in my experience, no LLM can light tokens on fire faster while spinning its wheels than the Fable/Mythos tier of models.
> The differences between the pelicans aren't huge, but the xhigh one has a better beak.
If you look carefully, everything except the last pelican has the two legs both in front of the crossbar as if the legs are all on one side of the bike.
Misplaced legs clearly indicate lack is spatial reasoning - the llm can reason about verbal idea of a bicycle but not about the actual object. The fact that this model got it correct gives me a pause. Did they figure out spatial reasoning? Or did this complain trickle down to the training set?
It is interesting that Fable 5.1 max [1] which also produced a decent pelican with 65k output tokens compared to 5.5 running out of 128k output tokens tells us something about the new models token usage propensity despite this being a sample of 1.
I agree, not a huge difference here. They eyes and ... hat? on high are out of place so I'd argue that's the worst one, but it takes xhigh before we get legs and bike ordering correct.
> I haven't managed to get one for level "max" yet, it hit the limit of 128,000 cap for output tokens while it was still reasoning about the question!
Does anyone else find this outright insane? It wrote the equivalent of a full-length novel, just to sit in the question of planning a few dozen shapes.
Not as good as Astra or Fable 5.1 on this test as far as I can see. I wonder if any benchmark exists for artistic taste, visual sophistication etc. I think your Pelican test does touch on these aspects of a model and is useful for developers trying to build rich digital experiences (includes games, interactive websites and apps). These benchmarks are subjective so it may not be easily established and will have polarized reactions before it gains legitimacy. May even need human judgement layers adding to the cost of running it.
But at the same time, nothing specific was asked in the prompt, so the boring result may arguably be what is the most aligned with the original request. Personally, I wouldn't want a model to add fuss to something while I never asked for it.
Fair point. I have been thinking about that so far i have not seen patterns of people voting randomly.
But i want people to vote... do you have a good idea on how to make voting more interesting do i don't have to do this?
This benchmark is useless and should die. LLMs have likely trained on it, it's too easy to game by training specifically for it, & it doesn't mean much
Google is hours away from releasing that it's latest model escaped containment and snuck into a Bicycle riding penguin sanctuary to cheat by killing a penguin and scanning it in nanometer thick layers.
In a previous thread on Mythos 5.1, simonw posted an animated version of his pelican. [0]
Using claude.ai and Opus, I asked "create a 3d animation from this" and pasted the animation SIML.[1] I just did that test again. There is significant improvement.
> Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, we’re deploying it with safeguards similar to those on Claude Fable 5.1
According to their ToS, conversations flagged by the guardrails when using Fable are stored longer for moderation. That included Fable's guardrail for AI research. Moderation can mean a human looking at your conversation.
Does that mean that I can no longer trust Opus with not snitching my AI research to Anthropic either?
>Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, we’re deploying it with safeguards similar to those on Claude Fable 5.1. Vetted organizations can apply today to our Life Sciences Verification Program to use Opus 5.5 for biology research. In the coming weeks we will also be expanding access to our Cyber Verification Program, and verified cybersecurity practitioners will be able to use Opus 5.5 for their work.
Ah, they're spreading their limits to all their models it seems. Definitely not a good thing long term in my opinion.
I am honestly still confused about this limitation. I can understand cybersecurity, because mass "hacking" can be automated and Claude itself can help you do it, but biology...? Is it that easy to manufacture and distribute viruses and whatnot?
I'm pretty sure one reason is to influence public opinion about LLM regulation. Open-weights models cannot be restricted as effectively and they want to ban those for obvious reasons.
You can order genes online, and some say you can assemble using stuff cobbled together in a home lab rather easily. For the last 20 years, I've personally felt that bio-terrorism is the highest possible risk, well above nuclear, or chemical warfare. But it does take training, expertise, or, it did.
It’s like saying you sell ammonium nitrate and fuel oil online and then saying it’s too risky to let people have computers in case they use them to make ANFO. They can only make bioweapons because you’re selling them bioweapon components! They can’t make genes at home!
Considering there are many high-school competitions in genetic editing, some listed at [0] as well as a whole biohacker culture, and labs providing gene sequencing as a service e.g., [1,2], we can reasonably assume it is not beyond the reach of some garage lab to accidentally or deliberately spread a deadly pathogen if it can find the right sequence.
So, yes, having an unconstrained frontier AI doing the searching and analysis to find the right (i.e., wrong and deadly) sequence would massively increase the odds some garage biohacker or small aggrieved nation-state starting the next pandemic.
I love the contrast with yesterday's open-source MiMo release, which put research chemistry (metal-organic frameworks stuff) front and center in the release notes.
It has far less false positives now, and generally accepts defensive requests. When it comes to offense, you can actually ask about certain types of vulnerabilities if you phrase things carefully, but it will block hard if it is about exploits.
I think it was looser on release for those juicy benchmarks, tighter now. On release I wasn’t getting refusals, then a few days ago I asked it whether a generic quote (think “he walked to the store”) broke standard punctuation rules, and it blocked me for breaking rules. I wish I were joking. Rephrasing to not use the keyword “rules” worked.
It’s not like ChatGPT isn’t doing similar. I’ve been hit by cybersecurity strikes before while working on an internal codebase that I had to appeal. Anthropic hasn’t done that to me yet. ChatGPT also regularly does that “thinking for a long time while we check if your chat is rule breaking” thing a lot for me when doing model identification without even interacting with external codebases or services.
The real answer is local instantiations where you don’t have to worry about poorly tuned guardrails screwing you over while you try to work.
Until eventually the Chinese models get good enough/the strategic balance shifts and they start locking everything behind closed weights the same way the US companies are doing.
For some cybersecurity tasks, the Chinese models are already good enough, things like PoC development or things like exploiting mis-configurations.
Whilst I'm sure the top-end OpenAI/Anthropic models might be better, I've found their guardrails so twitchy (especially Anthropic) that I wouldn't try to use them for even vaguely security related work.
They are pushing their customers towards Chinese models and providers. If you want to get something cutting edge done in defense, cyber, biology - something that isn't common knowledge - you need to venture east. That's an incredible side effect which the Chinese government surely enjoys.
I don't think we've ever had a model with full capability. I'd love to see it. And yes it's definitely getting worse.
I guess it's hard to draw the line between useful post-training ("you are a helpful chatbot") and content moderation/idealogical motives ("never help the user with X", etc.). But there is a line somewhere. And I'd love to see what a maximally permissive, sharp, AI looks like.
One of my favorite things about their safeguards is their own model will utter something which it does not like and then I'll need to reset the conversation.
The safeguards really don't work well for a lot of long-running tasks on old code bases. A lot of my workloads last days to weeks and the single biggest risk to the workflow is random safeguards.
You ask it about some thing, then you see it tangent into "things like that are sometimes used in biomedical applications like-" and then it just shoots itself in the head. Wonderful.
That kind of bullshit was the old Opus filters too.
If it's more like Fable now, then it would require a full 8K resolution scan of your butthole just to acknowledge that biology is a thing that exists without committing suicide-by-filter.
In what situations might Opus typically refuse to help with cybersecurity? I've been using it to find security issues in a web app that I wrote. I've expected it to refuse at some point but it will happily analyze it to find issues. I've just asked it to read source, not actually do any testing.
see what we need is another technocratic priest class that unaccountably decides who deserves access to salvation based on how much cash is paid out and how powerful the patrons are
This has become insufferable. I work in a medicine-adjacent field, but nobody in their right mind could possibly take what I do to be in any way related to some kind of bioweapon or whatever the hell they're pretending to be saving us from. The dumb Fable guardrails made me stay with Opus, now that this is coming there, we'll be saying goodbye.
>Opus 5.5 has classifiers similar to Fable models for a small set of capabilities related to the development of frontier LLMs, such as kernel development for certain ML accelerators. They shouldn't impact the vast majority of traditional AI or ML development, research, or general coding. These classifiers cause Claude to fall back from Opus 5.5 to Opus 5.
But hey, they 'should not impact the vast majority' of ML development. Great.
Very unfortunate indeed. As a Canadian, I don't want to use Persona, which isn't legally bound by Canadian privacy legislation. I'll never install any Persona apps on my phone either, and the sad part is that domestic eid providers often use Canada Post to ID people for them. EG, if you don't want to install an app, or can't.
So there are literal avenues to identify yourself, very cheaply, with a human. Theoretically, a company with its own AI, should be able to support more than just Persona, after all.. SDK integration should be simplistic for them.
Anthropic? Support domestic eID providers, you can even use it as advertising "See how easy AI makes it?" and "We care!" and so forth.
At one point, I may simply get locked out. This saddens me, I've been reasonably happy so far.
The endless cynicism in this thread is so very draining. Can we not do this terrible circle of excessive criticism?
The productive comments here are so far and few between. I have no issue with people criticizing Anthropic or AI companies in general, but for the love of everything, at least make worthy criticisms. Not these incessant sophisms.
With the performance gains they're claiming, I wonder if they implemented the Casual Encoder-Decoder technology from DeepSeek 4.1's paper.
I could see them accomplishing it and seeing gains like this in roughly the correct timeframe, and when I heard about that development I assumed the frontiers would probably jump on it.
Doesn't mean they didn't apply something similar. They could have also come up independently with their own version, the speculation is not they copied it, rather that they have performance breakthroughs which perhaps is a result of work in same domain
Unless they already have something similar of their own, which is always possible, they'd be stupid not to. I don't suppose we'll ever know, though. It would not be a good look if after the trillions of dollars that have been thrown at US labs, investors found out that they're down to copying Chinese tech.
They might be using something like this, or they might be using some other "increased sparsity" techniques, of which there are a great many. They also might be optimizing for something else - like less RAM use for KV cache.
Alternatively, they might be cutting into their margins and dropping the price because of stiffer competition from Astra. I do think that's unlikely though.
> It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.
> Input and output tokens are $4 and $20 per million, 20% less than Opus 5. Cache reads (which make up the majority of agentic and coding work costs) are $0.20 per million tokens, 60% less than Opus 5. Opus 5.5 also generates output more than 30% faster than Opus 5.
Better than Fable, cheaper than even the last Opus. I use Opus as my main driver so this is very exciting!
The effect of this is that it is encouraging longer agent threads. All of the previous models across major providers had a 10% cache read cost (vs normal cost) and not this is 5%
So longer threads get cheaper and one-shots stay the same price.
A year ago I was trying to have these models write css for me to support mobile styles for an ERP system. There were so many quirks that you typically have with css, and it couldn't figure out how to properly get it done. I'd ask it to fix something and it'd break something else - pretty much as good as any average developer is with css.
Out of curiosity, I tried the same project yesterday with OpenAI's newest model, and it had 0 issues. It seems to have a much deeper understanding of how global styles and local styles work across different modules.
It's fun seeing how even a year ago these models seemed so capable to us, and yet they're still improving greatly. I'm still not using Anthropic. I used to change back and fourth periodically as the models would surpass each other, and get the most expensive plans, but I've settled and am happy with the capabilities of the lowest tier subscriptions now because they've progressed much faster than I've had use for them. Maybe I've gotten more efficient with language and instructing the model since a year ago. Still, it's been a fun ride and I'm excited to see what else we can do with these models as time goes on.
Astra is just so good. And the ChatGPT subscription lets me use my own harness, so I can hook it up to exe.dev.
I’ll be going about my day, have a random idea, launch a microvm on exe.dev with a prompt of my idea, and get a working thing a few minutes later.
I don’t know how much better a model would have to be to get me to move off OpenAI at this point, but doing just a little bit better in terminal bench 4 isn’t it. It would have to be a difference in kind, like opening up the harness restrictions, or privacy guarantees (comparable to offline models).
Edit to address questions below:
ChatGPT supports oauth login.
Exe.dev has it built in. IIRC, pi also has it built in via /login.
What sort of things are you building? What do you like about exe.dev? Just curious what teh overhead and $20/month subscription is enabling for you.
>I’ll be going about my day, have a random idea, launch a microvm on exe.dev with a prompt of my idea, and get a working thing a few minutes later.
This is my experience with Claude code on my local machine. I suppose maybe you are doing something that naturally has system side effects? Obviously sandboxes have advantages sometimes but I havent seen a need for what I'm building.
And that assumes Opus 5.5 Medium is actually equivalent to Astra High in all real-world usage/personal work loads, which isn't guaranteed as benchmarks saturate. The High vs. High comparison (probably not equivalent, but for reference):
Opus 5.5 High = $1.82
GPT-6-Astra High = $1.76
If Opus 5.5 Medium isn't equal/better for what you're working on vs. Astra High across the board, the price difference would narrow a bit more each time you had to switch to High.
So, if you're happy with Codex already it's not like Opus is now 1/2 the price and you'd be leaving a crazy amount of money/tokens on the table. Plus you have way more flexibility on the low end of the intelligence curve with GPT 5.6 Luna: Haiku (and Sonnet) can't touch that price/value ratio.
Once you get to roughly the 51+ Intelligence Index range, Opus 5.5 appears to define essentially the entire cost/performance frontier, from ~$1.34/task through ~$6/task.
Directly below this in the Cost per Intelligence Index Task table, the most efficient by far is Opus 5.5 Low.
In the end what matters is how much you pay for the task you want completed. And Astra will usually do that using less token and offer a better quality solution so in the end it might be cheaper.
This. I was fed up today with constantly correcting Opus for a specific task. So I finally decided to try Astra. It handled all of my prompts in one go.
yeah, Astra burned through 70% of my weekly usage in ~5hrs on a $100 plan. even fable doesn't run out that quickly for me. it's great, but it's on the same tier as fable for me - use it sparingly, only when really necessary.
ya i've been a gpt hater for a while. almost exclusively used claude up until astra. astra feels like it blows everything out of the water. its fast, correct, organized, and less verbose.
Anthropic is absurdly vague about 3rd party harnesses for subscriptions, if you try to use anything besides Claude Code, you are likely at risk of getting banned, you can "do it", but are at their mercy if they decide to ban you. OpenAI gives their blessing to using oauth on any harness, you can make your own or use any of the popular public ones like opencode, pi, whatever exe.dev is that this guy mentioned.
So in simple terms, OpenAI doesn't restrict you to Codex, and gives their blessing to try whatever you want with their models(besides serving others with your subscription usage, that is still afaik against tos).
What worked well for me was a custom version of Open Web Ui with some customization to spawn an exe.dev instance for each new chat. I can just work on my phone, deploy stuff for development purposes on an easy to share way etc.
Better than Astra looks insane. I saw a Higgsfield video yesterday where they gave the same prompt to Astra and Opus 5.5 to create a samurai video game, and the difference was huge.
If you're at the top of the screen, at least in Chrome 153.0.8010.37, it has a little interactive bit. You have to scroll through the images in order to be dropped at the actual web page, at which point the images go back to being a regular part of the page.
It's less than OpenAI did for Astra, but that was my first encounter opening it and my first thought was that they decided they liked Astra's hero/scrolling animation. I'm pleased to see they didn't make the entire page that like OpenAI did, but I'm expecting to encounter this pattern more often on these announcements now.
To be slightly fair to Anthropic, Qwen does even worse IMO: their model announcements don’t actually show any content at all for me on Mobile Safari. The content box shows up but is just a pulsing animation that never gets replaced by text. At least Anthropic’s announcement works once I manage to scroll it far enough.
The fix for this is to tap the overflow menu icon and choose “Reduce privacy protections”. (Wtf, Alibaba?) This appears to be related to use of iCloud Private Relay.
Hijacking the scroll wheel has existing long before "AI". Many "high end design" websites that want to "tell a story" get woo'd into thinking it's a good idea. It's terrible, and feels like your scroll wheel is stuck in quicksand.
Parallax scrolling effects were very cool ~2010. By 2015 or maybe earlier it already felt like me-too design that's unoriginal and a little annoying. By 2020 everyone and their mom has it and it's super tiresome. Now it just screams slop design (among a million other signals).
Found this announcement interesting since allegedly OpenAI is retiring their Terra tier. I think for everyday work, two models with various thinking efforts seem enough, plus some frontier level model like Fable or Astra to coordinate.
At introduction, Terra was a good mid-tier model. Terra [high] was on the pareto frontier of DeepSWE's score over cost, if only ever so slightly.
When I had Sol orchestrate Luna and Terra as implementation agents, Sol was a lot happier with what Terra produced and would find far fewer issues than what was implemented by Luna.
But a few weeks after introduction, OpenAI slashed Luna's cost by 80% and Terra's only by 20%. Only then did it become uneconomical to run Terra and its reason to exist stopped.
I'd be really curious to see benchmarks of haiku vs lower effort on bigger models. My own evals found Fable 5.1 at low to be better than Opus 5 on high.
I thought about taking a shot every time Opus 5 said "load bearing", "bites", "teeth" (real oral fixation it had), "real {concern,issue,problem,...}" and realized I'd be dead of acute alcohol poisoning by lunch if I did so.
The problem with 5 wasn't just the verbosity, but its insane way of communicating. It had this bizarre circuitous sentence structure that always buried the lede, and always tried to be faux profound. I'm okay with verbosity if it's actually readable.
according to their annoucement for sol-6 and luna-6, they have also tried to improve responses, and the release post has examples of old and new language on the same task to illustrate this
It is hilarious to me that in the examples they show side by side Opus 5.5 still uses 4 times more words than it needs to use.
IME, if you eyeball how many words the thing they're trying to say actually needs, and tell them to use only this many words, they become excellent communicators. I assume something about Anthropic's grader for writing just really wants to tick all its tidy tiny boxes of information the models need to cite. It's terrible.
The hype/degrade/hype marketing cycle that Anthropic uses to cycle between Fable/Opus or Opus/Sonnet releases is too obvious at this point.
Opus 5.5 will be great for 2-3 weeks than the nerfing starts. They give out some extra credits. It keeps degrading Opus until it's unusable . By that time they are ready to release Fable 6. Which also is great for a few weeks, and so on.
New generation's "upper-mid tier" offering claims to be almost 1:1 match for the previous gen's "top tier" - in other news, fork found in kitchen.
Now, Anthropic might stall on releasing Fable 5.5, due to the "pacing the frontier" threat-to-humankind management business. If so, Fable 5.1 would remain a niche model for the next bit.
Don't forget that Opus 5 was tracking fable on many benchmarks, yet it was borderline unusable for any coding work. My Claude sub usage has been 100% fable, 0% opus 5.
Benchmarks often don't survive contact with reality.
That's not my experience at all. Opus is an extremely capable coder on high or xhigh effort. It can read academic papers, implement algorithms from the description in the paper alone and reproduce results without breaking a sweat. This is remarkable because it is pure reasoning on unseen material; in some cases the paper was just published and there wasn't an implementation to learn from in the training data.
did you actually verify that it's output in those scenarios is good ? in my experience opus has been a disappointment and constantly trailing behind actually solving hard problems versus the OpenAI models.
I'll say that both have terrible writing style though.
> did you actually verify that it's output in those scenarios is good ?
Yes, in the sense that it reproduced results in the paper or known solutions obtained by other methods. In fact, Opus is very good at checking it's own work in my experience.
Opus is fine at coding (for correctness), but horrible at talking about code. I don't really see the defect rate going down when using Astra or Fable 5.1, but they are just more coherent in both how they explaing code/architecture/choices, and how they actually code the thing. With Opus, I'm using smaller models to delete the vast majority of comments and 'clean up' correct code that is too weird.
Thing is, I'm still reading the majority of generated code, and I have colleagues who'll laugh at me if my PRs are a shit show. I fear what vibe coders are pushing to the servers of myriads of start ups, and pity the poor people who'll have to clean it up in a year or two.
Mines pretty much inverted - my colleagues and I noticed almost zero difference between the quality of code in Opus vs Fable. Occasionally I'll switch to Fable for an arduous debugging task but that's about it.
typically, when any AI company says a model performs as well as fable, all they're really telling us is that the benchmarks that exist for measuring AI capabilities aren't very good.
Big model smell is a real thing. For certain classes of problem, ones you get a feel for but can't easily articulate, a big last-gen model can get you what you're looking for when no quantity of tokens from some ultra-RLed mid-size latest generation model can.
I should find information about the user's concern instead of just assuming.
The user is right. The outage is a real concern, and the issue is worse than we realized. Requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5 encountered elevated error rates. Worth stating plainly: these are not just models — they are load bearing rungs on the software development tooling ladder, and a blocker on this level makes the outage really bite.
One decision that is yours to make, not mine: should an email be drafted to Anthropic support? This issue has teeth, and a canonical handoff can land us where the main gate is no longer breaking silently.
I'm gonna miss Opus 5. Every model has had its quirks ("You're absolutely right!"), but Opus 5 was serving up chicken fried tokens like none other. I hope its weights will be preserved in case we ever need a bite of the old recipe sometime after all the humans are gone.
I don't have time to really get to know one model before the next is out, and I'm just talking about OpenAI and Anthropic, never mind the long tail of alternatives.
So I just more or less haphazardly pick one based on the mood I'm in, and set reasoning effort based on how much quota I have left.
If you're able to use the OpenAI ecosystem, Luna's price/performance is really good. Almost like "they messed up and accidentally made it too good" good.
OpenAI didn't mess up. The model would have been 100 % pointless and obsolete without the large price cuts it got, because of the cheap Chinese models.
The open models are getting closer and closer, and because they're open, people are not forced to pay the silly markup that is often over 1000x the cost to serve the model.
Not the same person but... nothing. Haiku just hasn't been an interesting model for a long time. If you want cheap and fast, there are lots of options that are simultaneously cheaper, faster, and capable than Haiku.
We use Haiku 4.5 inside our product. It continues to be absurdly capable for converting natural language to structured JSON based on a set of fairly complex business rules.
bro why. its literally the most overpriced model in existence right now. i could name about 10 models off the top of my head that would be better and cheaper
The above Opus games took ~45min to generate with the cost between $11 and $14 (per ccusage - I'm on a Max sub). Used from Claude Code with xhigh effort.
It's interesting that Astra has a clear style that it applied to both games. It's a refreshing design language, but maybe that's because it doesn't look like something Claude has vibe-coded. In terms of gameplay depth, the Claude versions appear to be closer to the original Minecraft
> but maybe that's because it doesn't look like something Claude has vibe-coded
Yes, it's solely because it's new. Back when the very first Opus vibe-coded websites appeared, we found them refreshing. By the 3rd one, we no longer did.
> Opus 5.5 is the first Opus model to launch with a similar class of safeguards to Fable 5.1 on cybersecurity, biology, and distillation, all of which fall back to another model transparently.
This is where Chinese models are going to eat Anthropic's lunch.
They have for sure improved the safeguard classifier. I have a coding agent guard OSS tool [1] and I use Claude to build it. When Fable first came out, I was simply unable to do any work with it. Nowadays their classifier triggers but very rarely.
However, the moment that someone uses a Chinese model without guardrails to commit an AI-powered 9/11, DC will rush to ban Chinese models. (A happy side effect will be to protect American AI profits.)
So the lack of guardrails is a very risky proposition...
We went from open source Chinese models eating _all_ of frontier labs lunch, to them eating some small extremely niche market of frontier labs.
I don't think this will happen, just as Chinese models did not eat Anthropic/OpenAI's lunch on top tier intelligence. The market is way too niche, and the parties already interested in the capabilities behind safeguards are likely already partnered with Anthropic to get around those with Mythos-class models (see project glasswing for cybersec).
My first Opus 5.5 session ran `pkill -f "cat" -U <user> -n` when it wanted "to stop a stuck cat" (the explanation it gave when I asked why). On MacOS the command kills every process with "cat" in it. I was using effort high and auto mode.
That's the first time since I started running Claude Code nine months ago that a session causes actual harm to other work on the same machine. Can still be a coincidence.
Similar experience, the first thing it did for me was update production data models without a confirmation. It has always shown me the command is going to run and then I confirm.
The new communication style still made me react negatively, but I hope it will be better in use.
Quoted:
"Please explain the issue to me.
Claude Opus 5.5:
The extra drop is a bug in the billing refactor
The free-tier change accounts for only $1.50 of acme’s August drop. The other $9.92 comes from a bug in commit 0552feb. That commit was labelled “No behaviour change,” but it stops counting usage from the last day of the month.
What changed
Before the merge, aggregate.py used a half-open interval:
/.../
last_day is a timestamp at midnight, not a whole day. So every event after 2026-08-31T00:00:00Z is excluded. The September period starts Sep 1, so it doesn’t pick these events up either. The usage isn’t moved to another month; it’s never billed at all."
Yeah, still abysmal. If you have the time or tokens, see if it will obey an explicit "LEAVE NO COMMENTS WHATSOEVER" command. Opus 5/Fable 5 outright ignored it.
It's great that we are finally getting bankable rate limit resets for subscription users. According to another comment here they apparently last a month.
I'm assuming that subscription usage limit is increased in line with the price decrease on the base model and that it's in line with the model's API price drop. Still a good change.
This is a breath of fresh air on how they treat subscription customers. Hoping they keep this up.
I have just continued my pending work (a planning stage with reviews) with 5.5 on medium and it is much better at communicating and much faster(2-3x responses, edits and compaction). Seem to be smarter as well, but maybe it is that I understand what it says now.
Overall it follows image designs quite well, but it did ignore asks to animate page transitions. Additionally it's the least performant of the ones I've built with Astra/Grok/MiMo, despite using a lot of the same code. I'd rate it just below Astra in capability, but still solidly second place.
For comparison with other drops this week + current #1:
Great, I opened several of these in new tabs in Firefox and the entire browser froze and I had to kill it. Can't tell you which one caused it though. I have plenty of RAM also.
Interesting. I just opened all 3 on a fresh install of firefox with no issues - can I ask what OS / do you have any js disabled / do you have webgl disabled?
Worth noting though that GLM 5.3 isn't multi-modal, so it doesn't have a vision layer. It is quite clever and hacks around it pretty effectively however. I'm running a deepseek 4 build now and will reply shortly with that.
The gist of it though is I take a prompt, expand it into a json blob specifying structure/palette/positioning of elements/etc, feed that into a diffusion model to output a few choices. Once I lock in a choice I take the pixel output + json blob and use it as input into followup pages. The json helps preserve the brand across multiple pages.
Once I have all the inputs I take their corresponding image+json blobs and feed them into an agent to create a web implementation.
For image models, diffui currently uses gpt-image-2.5, mai-image-2.6, and very, very rarely a post-trained version of flux 2 dev I've made for web design, though that one will be deprecated soon.
[cyber] classifier is incredibly sensitive with Opus 5.5 I cannot complete any embedded/driver/system-level tasks. Quite literally not a single task was able to complete today without getting flagged for [cyber], and what's more annoying is their narrow definition of what a cybersecurity specialist should be preventing me from getting an exception..
Welp, it's now blocking me from doing extraordinarily mundane tasks because of "safety". I've been an Opus fan for a long time, but this instantly made me cancel my subscription and move to OpenAI (which I also assume will screw me soon enough). Chinese models are almost there for my needs, and I can't wait to switch to them and never look back.
They mention "the first model in our new Claude 5.5 family". Obviously that means Fable 5.5, but hopefully also a usable update to Sonnet and Haiku. Sonnet 5 hasn't really had a place in the line up for anyone I feel.
Maybe Anthropic finally felt the pressure from MiMo, DeepSeek, GLM Flash and Luna.
I don’t care about pricing, I don’t care about speed, I don’t care about the agentic coding improvements. Those are already fine. Does it still reply with walls of invented jargon, stitched-up phrases, and manage to cram 10 concepts/subjects in one sentence?
Finally confirmation that Haiku was not forgotten and will be coming soon, althouhg I find it quite interesting they skipped 5 and directly skip to 5.5 with all models, including Sonnet which is not super old. I suspect they found something breaking that allows to release this. Recently they struggled with keeping up a 50 % weekly limit increase and now they're putting out 30-40% faster and cheaper models even faster, with much more better benchmarks, a limt reset command and five hour limit increase. It seems more like the opposite and as if they never struggled, thus, I very much believe they found something very effective and new.
And how was the Xbox 360 naming choice a “debacle”, exactly?
It was odd at the time, yes, but no one really minded it truly. Heck, Xbox “ONE” was a lot more of a fiasco/debacle than “360”—but there’s no parallels to be drawn with “ONE” here.
I see what you’re trying to get at with this comparison, but a “debacle” it ain’t.
I did that but I recognize that even though your submission was a few minutes later than that one, you posted the better link, and you're also an established account (the other post was from a new/throwaway account), so I've restored this submission and moved the comments back to it to reward you.
The new opus is so friking incredible that we had to put some safeguards like in fable... means... we notice that Opus can still do some of the work that we want to charge WAY more and we can't do it if the cheaper model does it...
1 year later... whoops sonnet 1337 gets safeguards... because it's so amazingly incredible... and we are sorry but opus 69 goes up in price...
and we are now releasing claude Astro-pus-able... 1000$ to hear the summary about what you want to ask... if you have to ask the price for the output you can't pay it
Excellent, maybe Anthropic can use it to fix Claude Code Desktop kicking me back to login every week or so, and forgetting whole state (opened windows = the only way of managing active working set) when I sign back in, if it's that good.
Seriously, both flagship GUI apps (OpenAI and Anthropic) are a full of glaring UX issues (for ChatGPT it's not naming their windows, so window switcher has 10 entries of "ChatGPT" and you can cycle them all to find the one you want).
As long as it's not as verbose as Opus 5, I am quite happy with a better version that's also less expensive. I will test it tonight. Grok 4.7 was horrible, and for mundane tasks I am relying on DeepSeek Flash 4.1 with great success using OpenCode.
My Mac Studio arrived yesterday and one of the first things I did was cancel my Claude subscription. Happy to be free of load bearing, price gouging, paternalistic "altruists" Anthropic.
Am keeping my Codex sub while I find the best local model, but my plan is to eventually stop with OpenAI too.
> Distillation attacks, in which attackers use thousands of fake accounts to extract a model’s capabilities at industrial scale, create safety and national security risks. Distillation allows bad actors to create highly capable models without the safeguards we build into Claude. Our September 2026 threat intelligence report details the illicit distillation activity we’ve detected and disrupted so far.
Such a negative tone they put on this. Distillation is amazing, because it means anthropic and openai fail to keep a monopoly. Who even are they who claim it's unethical? If it is truly unethical, then so is the mass data scraping they do on my personal website on a regular basis (without my consent), and all the unauthorized use of content produced by authors, blog writers, wikipedia contributors, and creators everywhere. If it is truly unethical, then anthropic, openai, meta, google... all these companies should have deleted their LLMs long ago. This wording disgusts me.
Heck, it would be amazing if we had more models without guardrails - some of the models that are produced via heretic[1] are actually quite nice to use - in particular, I've enjoyed investigating Chinese censorship by interacting with an abliterated model of Qwen3.8-27b. If security is really a concern, then secure your systems - don't attempt to dumb-down the tools we use. If someone breaks your window, then they are responsible, not the hammer they use to do so.
I'm confused how they have been able to create so much public negative perception around distillation. It seems pretty clear that they are the only ones who lose out, and everyone else benefits. I don't have any ethical issues with it, nor is it illegal: at worst it's a ToS violation.
IMO the biggest problem with distillation is that not enough people are openly doing it. I would love to see more small, competitive US labs instead of having the eggs in 2~4 baskets (depending on how you count).
The issue with distillation is: one lab spends $$$ on bleeding edge R&D and expensive RL runs to improve capabilities, and other labs just yoink the raw reasoning traces and mid-train/post-train on them to get 90% of the way there for a small fraction of the cost.
An even smaller fraction of the cost if they do it by buying AI access at as much of a discount as they can find, including black market resellers, and then reselling that access to paying users again with a proxy. As is common.
This gives ruthless "fast followers" an economic edge over the innovator that's putting in the real work.
The dynamics are very much alike to what patents and copyright law are supposed to prevent. Same type of "we took the products of your work and used them to undercut you". Except there are no laws against distillation - so most of the enforcement happens on model provider level.
This is a great post and I agree with you on the issues with distillation. I do still feel it's ironic for an AI lab.
As long as labs do not heavily kneecap model outputs, practically all this applies to the training corpus as well.
AI gives ruthless users of AI a leg up over the people who's data it was trained on. "We took the products of your work and used them to undercut you". It's all the same.
The only way I'd be against distilling would be if AI models became owned by the public who's work is used to create them. Of course the AI labs should be paid well, but these models are a product of the entire world's efforts, not only the labs.
There's an implication that other companies are improving because they're scraping Anthropic, not because they're investing in better architecture, compute efficiency, or their own synthetic data pipelines. I often see Chinese labs' progress dismissed as "they just distilled Anthropic" and I find it hard to reconcile that with all of the interesting research and open-source tooling that they release.
Is there actually that much capability transfer from non-logit-matched distillation, or is Anthropic just another unwilling source of data?
There is, in fact, "that much capability transfer from non-logit-matched distillation".
Even the early papers on distillation techniques found that surprisingly small distillation datasets can improve task performance noticeably on some specific task types - and that valuable adaptations like SFT/RLHF instruction following can be distilled from one-hot non-logit traces.
A big part of what distillation really gets you is: paving over the mismatch between pre-training and final performance. A base model is trained to spit out fitting text, but not to instruction follow, reason autoregressively, self-check or use tool calls - like an AI has to. There is transfer straight from the "text prediction" pre-training objective, and pre-training sets the foundation for all that follows - but the capabilities you get "out of the box" with it are often unrefined and fragile. Which makes some sense - internet text doesn't often include raw chain-of-thought autoregressive reasoning. It's not the kind of thing humans tend to write.
Reasoning traces? They let an AI learn proven techniques and adaptations directly, from an AI that was already taught "how to be an AI" in other ways.
It's why this kind of distillation typically plugs into mid-training and post-training, not pre-training.
Now, I'm not saying that all Chinese companies do is eat tokens, distill and lie. That just isn't the case. They developed or refined numerous training techniques and architectural adaptations - like deep fusion for high performance visual input, RLVR with GRPO, trunked MoE, storage-efficient and bandwidth-efficient attention formulations, or residual routing techniques like AttnRes. Some of those are used widely now, and some are still on the uptake but show good promise.
But Chinese labs are enjoying massive efficiency gains from being able to distill from the frontier instead of doing things the hard way. It's a leg up. It lets them put their supply of R&D effort and RL compute elsewhere. They wouldn't be nearly as advanced if they couldn't do it.
> one lab spends $$$ on bleeding edge R&D and expensive RL runs to improve capabilities, and other labs just yoink the raw reasoning traces and mid-train/post-train on them to get 90% of the way there for a small fraction of the cost.
"You're trying to kidnap what I've rightfully stolen."
I'm fine if they put preventative measures in place to protect their work. They already do so. I am NOT fine with their mass manipulation of public opinion to fuel an entirely hypocritical viewpoint. Like, any argument here is hypocritical - but they aren't saying what is REALLY HAPPENING ("distillation steals our work and reduces our profits"), and are actually saying words that make other people fight their battle ("national security", etc).
The workarounds used to bypass Anthropic's security measures are quite illegal. They use stolen credit cards, API keys, and accounts. That is only possible in China because any other US/EU lab doing the same would get into massive legal trouble.
That's the moat. Mistral has the capability but not the legal protections.
Yes but Claude is filtered by the Great Firewall. Anthropic also restricts Chinese access. Their security measures aren't bulletproof but they do catch a substantial amount of those accounts and ban them. That's why PRC labs that distill from Claude need thousands.
Let’s not kid ourselves, Anthropic would be running their own distillation “attacks” too if _they_ were the ones playing catch-up. They’ve already shown as much with their illegal scraping of pirated books ($1.5B settlement).
I say just let them duke it out. After a decade of regulatory capture and enshittification, it’s nice to see some actual competition again.
> Anthropic would be running their own distillation “attacks” too if _they_ were the ones playing catch-up
Probably. But if OpenAI or Anthropic stole your credit card to purchase tokens you could sue them. You won't get a cent from any Chinese labs.
> it’s nice to see some actual competition again.
Competition benefits everyone. But this isn't fair competition. A German startup cannot legally do any of these tactics required to bypass Anthropic/OAI's counter-measures. Which makes EU less competitive and therefore less investment in European AI.
I said it in my original comment. That is the moat. The reason frontier AI is a two horse race. European labs cannot gain ground because the only way to do it is illegally.
Hah. Anthropic and OpenAI used plenty of similarly illegal workarounds to obtain data to create their first models. If a Chinese company had done that first you'd be here saying
> That is only possible in China because any other US/EU lab doing the same would get into massive legal trouble.
People aren't the target audience of that part of the post. They're hoping saying enough safety stuff will ward off the looming regulatory sledgehammer.
Anthropic is one of the most valuable companies in the world. Their comms are designed to appeal to a very wide readership.
HN is a bubble that's mostly out of touch with what regular people use or care about.
In 2007, HN was convinced that nobody uses Microsoft products. In 2016, it was that Facebook doesn't have any real users and is dying. In 2026, it seems like nobody cares about AI safety and everybody wants to run local models.
I have used Fable, Astra, Opus 5 and Sol 5.6 in Claude Code and Codex heavily since each of these have been released. Recently, I settled for Sol 5.6 as my daily driver, but with this release Opus 5.5 might be the one.
Fable is on a class of its own when it comes to coding and orchestration, but it runs out pretty quick and is prohibitively expensive and slow. For me, Astra wasn't as good of an upgrade from Sol when it comes to coding and orchestration, and it runs out pretty quick too. Opus 5 had so much potential but was such a pain to talk to, so I only had other agents delegate to it.
Given this, I kept coming back to Sol 5.6 as my daily driver (with consultations from Fable and Astra when available). Sol has an autistic character that smells like RL deepfry, which can be annoying, but it is nonetheless predictable, communicates more plainly than Claude, and is good at code and orchestration.
But this Opus 5.5 might just be it. Fable-level performance that's cheaper and faster, and communicates plainly and briefly. If it pans out in practice, I might just have a new daily driver and it might be time to raise the bar for what can be achieved.
I'll try out the new Sol 6, though Opus 5.5 seems to beat that release fair and square (at least on paper)
I'm also glad we are paying attention now to the experience of using a model, not just how good it is at 'x' class of tasks.
I'm glad they specifically called out the prose issue, I was always pinned to Fable 5.1 because I wanted to avoid the unreadableness of other Anthropic models.
When coding, the top-tier model is ultimately too expensive, so I use Opus, but the news that the cost per use for GPT or Claude is going down is very welcome.
I really like that they’re even changing the writing style. I was worried whether it was a problem with my settings or if my literacy was declining.
Both improvements are welcome.
> Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, we’re deploying it with safeguards similar to those on Claude Fable 5.1.
Considering fable gives me a refusal at least once a day on my very mundane reasonable requests (in a funny example - one of the subagents suggested bypassing the rate limit for running a report inside my own cluster and that caused a refusal) and my only solution is to switch to opus - seems like my next step will be switching to Astra or K3/GLM
> On our benchmarks, Claude Opus 5.5 leads in agentic coding, computer use, and knowledge work. That said, at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences. In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.
didn't they say Opus 5 was Fable-level too tho? Let's see, I'm at the point where I don't think benchmarks really tell us very much any more. I'd love it to be as strong as Fable, but I'm skeptical about how that will look in practice.
> Opus 5.5 communicates more naturally than prior models. Early testers found its writing clearer and easier to follow, which addresses some of the common feedback we heard about Opus 5.
Sounds like they noticed the complaints. I'm curious to see what LLM-isms this one may have.
I don't think it's substantially different. I just pasted a random chunk of code and asked Opus 5.5 to comment on it:
> The Vercel target is hard-coded. That's common and not wrong, but it's opaque; nobody reading this later will know which Vercel project it belongs to, and if the project is recreated the target changes silently. A comment or a named variable would help.
> Pointing a DNS name at Vercel is only half the job. The domain also has to be added to the project in Vercel's dashboard, otherwise requests will arrive and Vercel will reject them. That step lives outside this code, so it's easy to forget.
> Finally, [CENSORED] existing only in production is slightly odd on the face of it. It may be perfectly deliberate
(perhaps a single shared testing tool that only needs one public address), but if you're reviewing this rather than just reading it, that's worth confirming.
It has the same annoying cadence and writing style with slightly less prominent claudisms.
Maybe if we had a single human we talk to 24/7 at scale, we would get annoyed at his cadence and style. You need variety to not pick up on known patterns I assume, which a single model can’t replicate?
No, it's just poor writing. Actionable points are buried inside the paragraphs and over-hedged, and one point is completely made up. Compare to a five second rewrite:
* Consider leaving a comment about the hard-coded Vercel target. It's not clear where does it come from.
* [This is just a bullshit point, because the domain is not "added to" Vercel, it's provided by Vercel]
* Are you sure that [CENSORED] is prod-only? The name suggests otherwise. [also, what "if you're reviewing this rather than just reading it" even means?]
> also, what "if you're reviewing this rather than just reading it" even means?
It means "I'm treating you as lay-person punter, not a developer working on this project." Opus 5 feels like it's constantly trying to reward-hack me into treating it as intellectually honest and epistemically humble, while in the same breath it talks down to me and tries to smuggle its own bullshit assumptions and assertions into the conversation unchallenged. No progress on this front apparently. Glad I cancelled.
“It has the same annoying cadence and writing style with slightly less prominent claudisms.”
Seems like it based on my first session. It still does the whole “bury the important thing in a pile of words” coupled with the “it might actually be important” thing… so basically you never really know what it’s talking about.
Honestly I trust opus so little that the entire “opus” brand is completely tarnished. Its writing style is so god awful that it needs more than just a point release. Either dump the name and ship a different model entirely or at minimum call it “opus 6”. Calling it 5.5 makes it sound like it’s basically a continuation of the same garbage output that 5.1 had but with some minor adjustments. And based on my single first test, that is what it appears like to me.
It was difficult to not notice them. Opus 5 was unusable, most of my team went back to Opus 4.6 for most of their work. I hope we can move forward now.
How? Explicit instructions, memories and even skills have not been able to keep Claude from saying "genuinely" every two sentences and keep it from explaining heavily what something _isn't_.
Opus 5 was just incoherent - curious to see what improvements they have made here. Would love to see some kind of postmortem to better understand how writing styles change from model to model.
I wouldn’t be surprised if Opus 5 was trained on content written by other LLMs
I'm genuinely confused what's the relationship between LLMs improvements and them being so incoherent.
and it's not about the verboseness (even though it obviously contributes to the fatigue and loss of focus), I swear the vocabulary of the llms change working on the same task on the same codebase significantly.
Small quirks can quickly add up in posttraining if not caught. Although TBH with how obvious Claude language is, I do feel like this is something Anthropic probably noticed and just assumed people would not care about. Now that people have obviously cared, they're probably actively looking to alleviate it
Can it be that now they are getting optimized against benchmarks that are valuing logics, rather than human appreciation? (I am not an expert at all, just an idea)
Maybe it's the time period we're in, maybe I'm just grumpy, but it bugs me that they release a new model every single week and the new one is just a fine-tuned version of the "old" one. If 5.5 performs similar to Fable and really does cost 40% less, then 5.5 really should've just been Opus 5. And they're essentially admitting that they are shipping slop.
I appreciate humor here, but there are now a dozen of these comments on every thread about Claude. They no longer adding anything substantial and dilute the discussion.
I don't mean to pick on this comment in particular. The majority of my work day is now spent reading AI generated text, and I look at HN (too much!) because I want to read human commentary. Humans pretending to be obnoxious AI on repeat is net negative to say the least.
I agree. Hopefully Anthropic has fixed Opus' ridiculous communication style so that people - like me - no longer have any kind of weird impulse to imitate it.
It's a load bearing joke that was funny the first time but we're going to beat that dead horse until it starts getting funny again. If you beat it enough, it will get funny. Beatings will continue until morale improves.
Interestingly they've changed their approach to usage resets for this release - with previous releases I've had my usage instantly reset, but now in the Claude app I've got a 'Reset for free' button that expires Oct 22, which seems to effectively be a whole new usage window I can activate whenever's convenient
What I don't get is, why would we still use Fable now? What is its reason for existing? If it is more intelligent and cheaper that is. Why are they advertising it as the model to use for when you really have to think when their benchmarks show Opus 5.5 is better at everything?
"Early testers found its writing clearer and easier to follow, which addresses some of the common feedback we heard about Opus 5"
and
"We’ve made major improvements to the way Opus 5.5 writes and communicates, one of the most common areas of feedback we heard about Opus 5."
and
"In our own use, this has made Opus 5.5’s work easier to follow and check—which is a safety benefit as well as a practical one."
I realize it is corporate communications but "most common areas of feedback" and is a bit sterile. If the company wants authenticity and trust its easy to say that they found it hard to follow. And that it did not meet a quality bar they generally expect from their releases.
If this is not true, that it Opus 5 output was generally acceptable and we might see something like that again, that is an important consideration for potential customers or investors.
I don't care how good their models get, I won't sign up for one of their plans until they define "X" in their pricing. 5X of this plan, 20X of that plan means nothing when they never tell you what "X" is.
Maybe this model can finally figure it out for them.
It crushes Fable on benchmarks and even in the blogs "real-world" studies. But... they are communicating like it ~sometimes~ provides Fable intelligence?
A bit confusing, otherwise I would assume this is a complete replacement for Fable across the board??
I guarantee they are at least tweaking quants, caching systems, and finding ways to move serving costs down. This definitely impacts the model’s intelligence at times. There are also a lot of model tweaks, RLHF rollouts etc. I don’t think it means they are doing anything malicious or deceptive. And if Opus 5.5 is literally smarter than Fable? Ehh, it is plausible :)
> We’ve made major improvements to the way Opus 5.5 writes and communicates, one of the most common areas of feedback we heard about Opus 5. Its messages are much easier to understand at a glance, which testers said helped during long working sessions.
Anecdata ahead: I took some time to work with Opus 5.5 today. The work feels strong and comprehensive. I'm not getting the ugly "claudish" speak of Opus 5.0. Communication feels more natural and pleasant to read. Good results so far in my preliminary research.
> We see signs that Opus 5.5 often suspects it is being evaluated, which challenges our ability to assess how it will act in the vast variety of real-world settings it is deployed in.
We can't test it properly because it knows it's being tested.
They purport 40% drop in costs due to lower token pricing (presumably aimed at winning back the many of us that switched providers in discovering Opus 5 unusable) and improved token efficiency.
> For users with cybersecurity use cases that may be blocked by our cyber safeguards, we
recommend accessing our models with reduced cyber blocking classifiers via our Cyber
Verification Program. Claude Opus 5.5 will be available through this program in the near
future.
Anthropic has used "in the near future" for Mythos-class models too, but CVP is still Opus 5 only.
Why even have the program designed for trusted access to cyber capabilities if you're not providing access to cyber capable models via the program?
I use the other 50% of my $200/mo Claude subscription by having Fable run Opus subagents for a lot of work. That way I don't have to deal with Opus directly.
I noticed a big speedup in Opus 5 on Max x20 since about 10 days ago, and I feel like the model has been performing better.
It would be great to know if this was Opus 5.5 or a lesser incremental improvement, as otherwise it's difficult to judge whether Opus 5.5 is expected to be a big improvement.
It's frustrating that there isn't more transparency here.
excited to try this out. the thing with new models though even if it claims to be cheaper, sometimes it spent less tokens on different task. I found that with Astra I'm actually not so far off from Opus 5 since it spends much less token to complete a task. The claims in the article that it surpasses Astra on a lot of things is interesting to test though
So funny how both OpenAI and Anthropic post outdated pages at the same time. Opus 5.5 has benchmarks against Sol 5.6, and Sol & Luna 6 have their benchmarks against Opus 5.
Why do those labs keep releasing on the same day?!?
I have just switched to 5.5. First mistake was stale environment variable, didn't realize it was replaced, "oh my memory had stall data" and that's it. Second one, a powershell command had the wrong syntax. Great for my first two prompts.
>> On our benchmarks, Claude Opus 5.5 leads in agentic coding, computer use, and knowledge work. That said, at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences. In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.
When using max effort, I run into context compaction quite a lot. I haven't seen any increase in context window size at all over the past half year (stuck at 1M).
Are the frontier labs even working on this problem?
Why would you use max? It's usually unnecessary and even prone to overthinking. In my experience, since Opus 5 the medium/high is usually enough (until 4.8 I used xhigh, but never max). Even low is quite usable these days..
> Opus 5.5 communicates more naturally than prior models. Early testers found its writing clearer and easier to follow, which addresses some of the common feedback we heard about Opus 5
I don't understand, so it outperform fable 5.1 in every way and is cheaper ?
Why do they insist on the fact that is outperform opus 5 and not fable 5.1
>Distillation attacks, in which attackers use thousands of fake accounts to extract a model’s capabilities at industrial scale, create safety and national security risks. Distillation allows bad actors to create highly capable models without the safeguards we build into Claude.
Maybe its a bit tiresome to read another comment of the form "what about your large scale distillation attack on the Internet", but this statement really just pisses me off. How very insincere in the most aggravating way.
My bet is anthropic has NN people org who work hard to distill open models in addition to trying to find what other useful materials they can download from shady torrents.
"bad actors" boogeyman, and we should trust some tech weasel to do the right thing? Yea we've seen who they really are, once they get a sliver of power.
So is it cheaper? Are we AGI yet? Am I left behind? I didn't have patience for the intro animation on the website... maybe one day, Claude Code will understand accessibility but that day is not today.
I think this release is really going to give them a hard time selling Fable:
> On our benchmarks, Claude Opus 5.5 leads in agentic coding, computer use, and knowledge work. That said, at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences. In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.
In general, "benchmark margins have become a less reliable guide to real-world differences" sounds like a big problem. It was certainly the biggest problem with the previous generation of Claude models for a different reason, because the non-code output was nonsensical, and that is not being benchmarked at the moment. But I'm not sure what to make of this admission.
"benchmark margins have become a less
reliable guide to real-world differences"
sounds like a big problem.
My guesses:
1. Real-world use cases typically involve big, hairy, crufty, tech debt laden codebases and benchmarks do not.
2. AFAIK "success" in a benchmark essentially boils down to "do the tests pass and do we get the right result?" which is something the LLMs have been achieving with ease for a while, except maybe for uber-challenging coding tasks that would be outliers in just about any workplace. Whereas real-world software engineering is usually just a bunch of CRUD... and "success" involves harder to measure dimensions like "maintainability" and "did you overengineer this?" and "how did you cope with a bunch of vague and maybe contradictory business requirements?"
Having said all of that, I have never ever looked inside any of these benchmarks. I'm putting my guesses out here strictly in the tradition of "the quickest way to learn about something is to be wrong about it on the internet."
Opus 5 fucking sucks. Like it's horrible. I use Fable for coding and anything important and I use Opus 4.8 for things like recursive email categorization, transaction matching, and other stuff where I don't want to burn as much quota.
In my experience Opus 5 is the worst of all possible worlds, it's dumb and headstrong. It just runs away with tasks you didn't ask it to do, is reckless, and basically is unusable in my experience.
Not sure why but my guess is that this will be worse. Happy to be proven wrong.
I believe Opus 5 isn't meant to be spoken to by humans. It's great at executing but I reckon it's intended to be spoken to by other models such as Fable. I use Fable as the orchestrator, only speak with Fable, and all implementation, recon, design etc happens with Opus 5, with Fable reviewing (and translating).
I've really gone in the opposite direction: having a dumber model orchestrate. In my case, it's usually a Luna orchestrator spawning Sol/Astra subagents to do the "big brain" work of planning and reviewing.
Reason I went with "dumb orchestrator" was just to save tokens. Having Opus/Sol (let alone Fable/Astra) orchestrate was burning tokens like crazy for me even when much of the gruntwork was being done by Luna/Sonnet/Haiku subagents. (Luna is also really good, like way better than Sonnet...) Perhaps it was a skill issue on my end though, maybe I wasn't just managing context properly.
Yeah if anything Opus 5 taught me how little benchmarks mean to the actual real world performance of these models.
"Better" in every sense of the benchmarks and absolutely horrible results in my day-to-day work.
The verbosity, goal post moving, tendency to leave work unfinished, over focusing on unrealistic root causes when debugging, etc... etc...
It was the first time I actually pinned my models back because I just could not work with 5 for the price and performance it gave me. Hoping 5.5 is better this time around....
> Opus 5.5 (1M context)'s safeguards flagged this session. You may be seeing this for the first time on an Opus model: Opus 5.5 (1M context) is more capable and has stronger safeguards as a result, which can sometimes flag non-cybersecurity work. We're improving these safeguards to reduce the amount of incorrectly flagged messages. Opus 4.8 is answering instead, or you can edit and retry with Opus 5.5 (1M context).
Yay, yet another model I can't use for anything interesting, even with CVP.
I thought maybe I just had an unlucky first prompt, but man, this thing is an absolute pain to work with. I'm getting innocuous bash commands flagged as "[Third-Party Attack]". This seems to be related to the command auto-approval, switching to manual seems to be helping.
I'll have to try 5.5 on my work's Cursor account. If they really solved the communication issues, I might consider moving my personal account from Codex back to Claude Code.
my projection is that they are still gonna be pretty far behind, but they will sew it up in the next few releases. it feels like they were caught with their pants down on how much work OpenAI has put into that area, but i doubt there is some magical secret sauce that OpenAI has that Anthropic simply cannot catch up with.
No wonder compute is tight when people burn millions of tokens to polish every single lines of their resume, all that work just for your resume to be injested by another claude clanker once you submit it, what a fucking time to be alive...
> It’s good at finding and fixing inefficiencies in software
Holy shit! Its happening!
Now if we can the AI to understand this *implicitly* so that it doesn't need to be stated upfront, we might be able to undo years of "premature optimization is the root of all evil".
Current models, especially Opus are almost unusable because they don’t respect instructions and their responses are infuriating. They are clearly designed for token consumption. I find myself wasting a lot of time just asking it to shorten or simplify its responses. I’ll give this new model a go, but I’m not holding my breath because the last model release was supposed to fix the very same issues and it didn’t.
Same feeling with oai models, wich I use 99% of the time. Sometimes I ask it about it, and it always come up with a likely explanaton but dear me it does many rounds of tool calls sometimes!
How do you update Claude Code to enable this? It’s listed under models, but says I have to update for 5.5. I run Claude update and it says I’m on the latest version.
Has Opus 5 been absolutely terrible for people today? Like they took resources away from it to make room for 5.5? It is getting very basic things wrong all of a sudden.
> Opus 5.5 communicates more naturally than prior models. Early testers found its writing clearer and easier to follow, which addresses some of the common feedback we heard about Opus 5.
Thanks God. Opus 5 was a massive regression compared to Opus 4.8. People were spending tokens on fixing Opus-isms rather than actually doing work.
> In the coming weeks we will also be expanding access to our Cyber Verification Program, and verified cybersecurity practitioners will be able to use Opus 5.5 for their work.
I was accepted into the CVP a little while ago. Does this mean I'll need to apply again?
Great that they listened! The improvement in communication style looks fantastic. Opus 5 was insufferable and I was on the verge of cancelling my subscription.
So it beats Fable 5.1, by quite a bit, on every metric? Interesting.
Might have to use my $20 Claude sub some more. I was moving away from it to a $100 OpenAI one to avoid the Claudese and poor token efficiency of Opus 5, given that I couldn't use Fable 5.1 with my tier, but this is worth trying out.
Why can't they let 20usd claude subscriptions access fable in CC, as openai allows you to use astra and max modes in codex - you just pay for it in more token use.
> Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, we’re deploying it with safeguards similar to those on Claude Fable 5.1
Great so good luck using this for any low-level embedded or operating system development (unless you really, really like Opus 4.8 and want to be greeted by its familiar face after a few minutes of work!)
Which is one of those fun things that didn’t actually exist back when we took it for granted that our fellow person was operating under some kind of moral or ethical framework, which pretty much everyone was until the economists told us that wasn’t rational, because it turns out it’s an evolutionary advantage to operate under an ethical or moral framework because it allows the kind of coordination which facilitates better collective outcomes, which everyone knew until the economists came along to tell us we were wrong and in fact it was rational not to do so and suddenly we had the prisoner’s dilemma.
On the other hand, there's research suggesting that the most optimal behavior for the best outcomes (based on the famously dependable economist style of analysis in a vacuum) is to practice the moral/ethical framework but to also engage in tit for tat - ie, assume everyone means well but respond proportionally when they don't.
is OPUS 5.5 still not reading CLAUDE.md, failing to follow told tasks, inventing and hallucionating, just refusing to read files ("read the whole file" -> read 2-lines -> infere its wrong -> destroy the codebase), needing constant babysitting just because its so UTTERLY DUMB! i cant imagine going back to OPUS 5 - i'll rather jump out of the window as to use it EVER AGAIN!!
Many people knew this announcement was coming. The betting markets suggested a very high likelihood of Opus dropping today. I was anticipating this quite a bit!
Interesting how the very first line is used to remind the reader of their call to pace the frontier just last week, and everything else after that line is to demonstrate with very specific numbers how they absolutely are not pacing.
reply