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
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
It does work out to be a similar cost per task though
https://artificialanalysis.ai/models/claude-opus-5-5#intelli...
It is most of the pareto frontier.
https://artificialanalysis.ai/models/claude-opus-5-5?models=...
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.
Competition works.
If 5.5 is any better, I might try to do agentic-assisted development instead of just telling fable to delegate
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.
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.
> 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.
For long running tasks it is. That's what made Deepseek so cheap.
It doesn't look like that's happening, on the contrary the prices are falling especially when taking into account capabilities.
I'm hardly a fan of China/Xi, but I do appreciate and benefit from this.
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.
> 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.
> Rewrites are declared by the publisher, never inferred from overlap
> NULL means dirty, and DELETE is the fence
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.
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.
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.
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.mdIsn't this basically the model admitting it was trained on this? Otherwise why would it think a pelican svg is a usual request?
"Ah, yes. This is a classic dog-breed-to-appliance-failure mapping problem."
But its safe to say that pelicans on bicycles are disproportionally huge part of their training data
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.
Opus 5.5 High: https://threejseval.com/models/claude-opus-5-5-high
You can compare any other model on the same prompt. Gallery unlocks after 4 votes: https://threejseval.com
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
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.
The last pelican gets this correct.
Ah, they're spreading their limits to all their models it seems. Definitely not a good thing long term in my opinion.
ChatGPT 6 Pro answered it without issue.
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.
[0] https://www.sciencebuddies.org/projects-lessons-activities/g...
[1] https://www.genewiz.com/public/services/sanger-sequencing
https://mimo.xiaomi.com/mimo-v2-6#co-scientist-for-materials...
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.
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.
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.
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.
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.
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.
How it works: https://miraflow.ai/blog/deepseek-v4-1-flash-causal-encoder-...
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.
tired: AI startup attempting to build their own website
wired: a nonprofit founded in 1996
> 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!
So longer threads get cheaper and one-shots stay the same price.
It's just a standard hero image + text for me, with no scrolling effects.
edit: @iAMkenough figured it out, it was because I have prefers-reduced-motion enabled.
I agree that it's sort of stupid, not a fan.
For a marketing page, it’s not the worst UX I’ve seen, but still slightly annoying.
Everyone that doesn't gets served some animated bullshit.
God I hope so
Nice. I was starting to think that Haiku got abandoned.
I would maybe use Haiku 5.5 for highly parallel workflows like checking in on MRs or scanning my entire codebase.
All of this starts to feel more like a drug dealer selling their newest stuff.
In two weeks we probaly get Fable 5.2 with “groundbreaking” improvements, then Astra x+1 etc and then the cycle starts again.
And on the way I always have to check my tooling and need to adjust things to get max results.
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.
Yeah, like Apple tells me the M6 is the best chip, but just a few months ago that's what they said about the M5. What a bunch of frauds.
Benchmarks often don't survive contact with reality.
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.
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.
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.
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.
https://artificialanalysis.ai/models/releases/claude-opus-5-...
Opus 5.5 Medium = $1.34
GPT-6-Astra High = $1.76 Opus 5.5 High = $1.82
GPT-6-Astra High = $1.76So, 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.
The Claude lock-in simply disqualifies anthropic entirely (for my use).
Can you give more details here? This sounds intriguing.
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).
This is news to me. Excited to try it out! Thanks.
Terminal-Bench 4.0 - Stanford & Laude Institute (with funding from all of the AI companies)
FrontierCode v1.1 - Cognition
CursorBench - Cursor (now SolarBoringSpaceXAI I believe)
GDPVal-AA - Artificial Analysis
AutomationBench - Zapier
Humanity's Last Exam - CAIS and Scale AI
Terminal-Bench-Science - Stanford, Laude, Ai2, Allen Institute
OSWOrld - XLANG Lab @ the University of Hong Kong
Chartography - Surge AI
This is where Chinese models are going to eat Anthropic's lunch.
About the time.
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.
Maybe Anthropic finally felt the pressure from MiMo, DeepSeek, GLM Flash and Luna.
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).
Design: https://image.non.io/78795662-8bfc-4e14-8d72-3738392aa6b3.we...
Opus 5.5's output: https://html.non.io/annui-opus/
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:
Astra: https://html.non.io/annui/
MiMo: https://html.non.io/annui-mimo/
Grok 4.7: https://html.non.io/Annui-grok/
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.
Are the frontier labs even working on this problem?
Is the Xbox 360 (Xbox 2) vs PS3 debacle all over again.
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.
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.
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.
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).
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.
Is there actually that much capability transfer from non-logit-matched distillation, or is Anthropic just another unwilling source of data?
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.
That's the moat. Mistral has the capability but not the legal protections.
Instead of instilling confidence, it was overwhelming. Not sure if I'm the only one.
Has oneshot all of the quite complex bugs / debugging tasks I gave to it which I know opus 5.0 would've struggled with
They write that at the top, but then on benchmarks, it beats literally every other model, including Fable and Astra?
Will be interesting to see how people's opinions of it line up IRL, but so far I've loved Fable so hopefully will love this one too
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.
Sounds like they noticed the complaints. I'm curious to see what LLM-isms this one may have.
> 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.
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.
* 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?]
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.
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 wouldn’t be surprised if Opus 5 was trained on content written by other LLMs
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.
I wonder if there are studies around this.
https://openai.com/index/where-the-goblins-came-from/
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
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
"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.
> Reset for free: Get extra wiggle room to explore Opus 5.5. Expires Oct 22.
Wdym Opus 5.5 scores 14.7% higher than GPT Astra for Terminal Bench 4.0?
How would this alleged difference (most likely bs) actually show up in reality?
GPT Astra was literally the best model in the world by a margin until 1 hour ago or so.
>> 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.
> how would this alleged difference (most likely bs) actually show up in reality?
Furthermore: so they admit it's bs but still placate it like its the next biggest thing ever ... alright
All I'm saying is I refuse to buy into it anymore – yet many on here still do, including ... you?
A bit confusing, otherwise I would assume this is a complete replacement for Fable across the board??
Thank you.
It does perform slightly worse than Opus 5, but it is significantly cheaper and faster.
Opus 5.5 (med, as it's better than F5.1 high per graph in the article) used $2.2 and caught errors that Fable 5.1 missed.
Try Opus 5.5, cheaper, faster, and more intelligent for those prepping for interviews.
---
I provided crapton of context for that one resume line. All the work I did, documentations for my justifications, etc.
I initially messed up and came out ot $5, rest of resume used around $4 per line (I used a fresh new session on purpose).
---
As a clarification, $2.2 average for OPUS 5.5 was the same process in a new session, same context, same prompts.
Also adding verification for that Fable 5.1 output in the same sesssion.
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?
> 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
Maybe this model can finally figure it out for them.
Not efficiency in writing, clearly.
Yay, yet another model I can't use for anything interesting, even with CVP.
We can't test it properly because it knows it's being tested.
> 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.
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."
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'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.
"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....
Nice. I was starting to think Haiku was going to be abandoned.
Ants: It's a good model, sir!
I tried Opus 5 and Astra.
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.
https://www.reddit.com/r/codex/comments/1wnggya/gpt_6_droppe...
Resets Get extra wiggle room to explore Opus 5.5. Expires Oct 22.
What the hell does this mean? There are weekly "resets" anyways. And there will be 4 of them before Oct 22.
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".
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.
I was accepted into the CVP a little while ago. Does this mean I'll need to apply again?
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.
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!)
Less companies involved means less pressure to go fast.
chinese models can't come soon enough
we're already getting enshittification
Infomercial at its best.
No wonder we are hammered with ai announcements.
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.
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.
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.
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.
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.
You are all getting mad about absolutely the dumbest thing when there are giant things to be worried about here.
eg "pacing the frontier" could also mean they are impatiently or anxiously walking up and down the border.
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.
Is this the meaning or do I have it wrong? I have not checked.
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
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.
but without using the word "regulate" which is a negative connotation to business
but a "pacer" would be a leader of a pack which is a positive spin
it's classical business marketing language silliness
A world exists beyond your vocabulary, post it. Apparently, quite a big world.
It's a strategy to achieve more, not less.
A pacer in a race runs at a steady, predetermined speed to help their runner run at a target pace.
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.
It's a weird phrase. Not sure why there are so many people who feel the need to defend it with such passion.
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.
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
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.
Wtf is the meaning? Means absolutely nothing to me having not seen the apparent announcement last week introducing the obscure term.
sir, this is a hacker news thread
"there is nothing outside the text" - Jacques Derrida
https://www.war.gov/News/News-Stories/Article/Article/264106...
if you are in a long race, you don't run all out teh entire time. you pace yourself.
https://en.wikipedia.org/wiki/Pacing_strategies_in_track_and...
That couldn't be more exactly what they are doing here.
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.
Claude says it sounds fine. And Claude is now the judge of the English language style, not you.
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.
Do you have a better proposed phrase?
"Pace yourself" specifically means "slow down".
But stating it plainly like this would make the contradiction too obvious.
Though tbf corporate-speak and AI-slop are both insufferable in similar ways...
And I don’t think any pacing is/was intentional. They’de release skynet if they could and the stonks went up
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.
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.
Its an agenda serving blog post, but constantly bringing it up like this is just obnoxious.
> 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".
https://en.wikipedia.org/wiki/Occam%27s_razor
doesn't sound like a razor at all
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.
that, they fully intend to 'pace'.
what anthropic have stolen they intend to keep for themselves.
If their scare was honest, they would stop.
Pacing is very explicitly about RSI and similar training methods that will accelerate progress beyond our ability to comprehend it.
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.
Simply make them something that derives a text response from its training data.
Translation: our models are getting shittier each iteration and we ran out of ideas. Let's invent scary stories and hope investors will lap it up.
Idiotic.
The only good news is that these models are genuinely helpful and we have competition at least between 2 companies.