Beyond all the benchmarks, I think Fable 5.1 is a big improvement in writing style. It sounds a lot less stereotypically like other Claude models, has (imho) a much more natural style, and responds to my style instructions more reliably. More work to be done (and we will!) but reading better prose makes me so much happier.
Another point I expect not to get much attention until it all happens at once is science. People have been correctly excited about the many "sudden" breakthroughs LLMs are making in Maths, but some of the science benchmarks make me believe we'll soon see similar developments in other scientific domains. Fable 5.1 more than doubled Fable 5's Terminal-Bench-Science [1] score, which I think is meaningful.
I have a pet theory that the Opus prose style/smell we all have grown weary of is due at least in part to the models writing more for themselves and each other than for humans. They're packing lots of signal into fewer words and they don't care if it sounds cringe because it works better as glue in long-running tasks.
I'm also thinking of the 2017 novel "Void Star" where AIs who operate everything have long since left ceased bothering with human languages, and it takes a rare sort of direct matrix-gazing savant to be able to try and horse-whisper them into doing or revealing anything they didn't already plan to do.
There's a huge difference between the kind of prose you see in final output vs CoT windows. The final output is very much not what I'd call "packing lots of signal into fewer words" (aside perhaps from "Claude-isms" being easy enough to scan for if for some reason you actually wanted to scan for them, which other agents might want to for all I know); and if agents are writing for each other then presumably they could stick to CoT-speak (unless it's a distillation risk?).
Yeah, if anything the problem is that the output uses too many words for too little signal, and incorrectly uses confidence based on insufficient information to the degree it’s clearly bullshitting.
Our current AIs would do this now except there is a lot of human pushback in training because of interpretability. Otherwise it's just an emergent behavior that models will encode shorter token strings to complex concepts because it saves tokens/compute when running making the system more efficient (supertokens).
Of course these supertokens or other forms of language compression when you have a different model making sure the system is aligned and reads "red_ball bounce calcium" not realizing it means "grind the humans bones to dust" can be problematic.
This post and comment makes me believe "science" is the new "code" for Anthropic now that the code advantage is mostly gone and lost for OpenAI, ie. they got much better and Claude become significantly worse over these months.
IMO, Codex is worse than Claude with Fable. At least at Rust.
That said, the open source models are not bad and I'm looking forward to more tools and products built on top of them. Code review, security review, etc.
Anthropic needs to change how it treats users though. I'm increasingly put off by Dario, the rug pulling, the lies, and the attempts to regulate open weights. I'm going to bail if this doesn't change. There's plenty enough that's good enough, and those things are hackable and extensible.
If Fable isn't available at subscription price via third party harnesses soon, I'm also going to bail.
Its "pure capabilities" are definitely worse than Fable, but I find codex has a much more pleasant style and is in comparison much more generous with its limits.
Ugh, people are still saying the Codex limits are more generous. They're not, Claude's are over 2x higher! [1] It's just that Claude uses far more tokens, 2-3x is typical. Except sometimes GPT will use just as many or even go into a compact loop and then your quota is exhausted.
Maybe it depends on the type of work you do, because for me it almost never happens.
>> You can be 95% complete with the plan for it to trip and then lose it all.
That's... not what happens though. The session will either seamlessly downgrade to another model mid-session, or it will stop with an alert and you can just re-prompt it. It will still have access to the context.
As a fervent Claude Code user who made the switch to GPT 5.6 Sol over Opus 5 over hard-to-read prose this makes me happy. I love your product but the current models are very hard to work with if you need to do a lot of context switching. Brevity is key.
Which is something the providers that are trying to watermark their texts can't afford. Superfluous replies give much more opportunity to further encode this junk information.
Also a codex user but for me brevity is not it's strong suit. I basically have to give it bigger tasks than I am used to to warrant the time it takes to complete. I feel whatever context the tooling adds can also be problematic
Lets see what they do with Opus first. I didn't find Fable 5.0 prose that bad to read, but improvement is always welcome. It's Opus 5.0 that's atrocious.
How is it possible that all models from xAI, OpenAI, Anthropic, Qwen etc. win all benchmarks on each release?
Tomorrow all of the above (except Anthropic of course) will bump version numbers and be at the top of HN winning all benchmarks.
Science breakthroughs incoming? First of all, you are already restricting science in Fable, secondly, we have been hearing the same for several years now.
Please bring to the other models, and also please only apply the AI text watermarking only to EU citizens. I may not be able to tell when Claude writes about things i don't know, but in CC it writes about my code and it is obvious.
"this watermark is invisible to anyone who does not have the detection API"
1. This is BS since i can detect it when it writes about my codebase
2. I do not want secret codes being written inside my codebase, or anyone else's codebase that i use. The constraints of how to code why eliminate it from code itself... but there is a lot riding on the word "may". And even if it is just comments, this might explain Claude's desire to write such long ones -- long enough to encode secret messages in out material.
You’re probably better off organizing a campaign to pressure Congress to prohibit American corporations imposing foreign laws on Americans, which is what this text watermarking is, regardless of how you feel about it. I think it’s a precedent we really don’t want to go down if you believe in democracy and self-determination.
It also clearly establishes or the very least moves in the direction that you don’t actually own or control the output of AI in any manner whatsoever, you’re just paying for it since Anthropic in this case can simply essentially brand/tag all your output that is based on not directly your own words, but a higher level process or methods that you use, including your instructions and how you structure your information and what your overall objective and goal is.
Anthropic is branding it on the behest of the EU lew, which already is an entity that is diametrically opposed to democracy and self-determination based on its structure even if you ignore the fact that it violates the most fundamental concepts of self-determination in its direct contradiction of the UN Charter and implicitly the Universal Declaration of Human rights.
What people done seem to be catching onto is that the EU is becoming the world dictatorship because the USA has simply had too many onerous people and that stupid constitution and its amendments that keep roadblocks world domination for the ruling class vampire.
I think Congresspeople hearing that EU AI Act is forcing secret codes into the infrastructure of American technology across all industries is sufficient.
I'm really glad for that! And I appreciate that you're making yourself available. I really do. Outreach is amazing. And thanks for making Claude.
I really do love Claude. In some ways, I'm asking this question because of just how much I am grateful for the role Claude has played in my life.
> Fable 5.1 more than doubled Fable 5's Terminal-Bench-Science [1] score, which I think is meaningful.
But my honest question is, can I use Fable like that? Can I use Fable to do science?
To borrow a Claude-ism, this is "load-bearing" because Claude's response has been degraded for innocuous research projects concerning population-level analyses of astronaut health.
These "safety filters" trigger on questions about rabbit sex, smartphone accelerometer data to classify cat purrs, and so much more. What exactly does this score mean for users like me if it's unusable for middle school physics, biology and chemistry?
Second, I would happily quantify it for y'all, but qualitatively it feels like Fable's performance is noticeably poorer than initial release / launch.
And I am wondering if this is the case particularly for me because I use Claude via Claude Code to make a personalized care dashboard for my doctors to help me in managing my care.
"In the case of Fable 5, when a classifier fires, the model re-routes the user’s request to Opus 5, a capable model that does not have the same level of biological capability as Fable 5 and which cannot provide as much assistance to a malicious user. This is the fallback that users see when their requests are blocked."
I hope that I'm off base here, but I noticed that the post avoids saying that the user is informed every time when such re-routing occurs. Would you be open to confirming whether or not this is the case?
Is the end user informed every time their query is re-routed?
Or, can you confirm that there aren't scenarios where a user's outputs are degraded without telling them? I recall that this was something that had been adopted as policy for AI research during Fable's launch.
I sincerely hope that covert response degradation is no longer practised as policy.
Sorry for putting you on the spot, but again, as Claude would say, it's because Claude's load-bearing in my life. ;)
Too bad. I see the stereotypical prose as a good thing. When I interact with Claude myself, I don’t mind it as it just feels like Claude’s distinctive voice. But when other people try to disguise LLM output as their own thoughts, the voice makes it easier for me to tell.
People that want to be open about the source of their text will just tell you where it came from.
People that want to obscure the source of their text would rather that it was more difficult to sniff out LLM-generated text. And they're the ones picking which model to use.
> More work to be done (and we will!) but reading better prose makes me so much happier.
I assume this work will be done for Opus as well? Opus has seemingly gotten progressively worse at its prose and technical writing with each version. I've stopped using Claude entirely for now, because it manages to turn even the simplest technical explanation into the most obtuse and obfuscated word salad imaginable. People originally adopted Claude because it felt pleasant to use in comparison to ChatGPT, but I feel like that's really been lost (at least with the Opus line).
It's gotten to the point that I feel a sinking feeling of dread the moment I see a wall of text generated by Opus. Every developer I've talked to feels similarly right now.
Yeah, it's like day and night. It used to be really unpleasant to interact with early codex versions. Even 5.3 wasn't great. Now, I go to Sol if I need to discuss anything. I don't even bother with Opus because I know that it's going to give me a headache.
Well your CEO went on X saying you will cure cancer, and since it's always a 6 month rolling window with him I can only assume humanity will be cancer free before next summer, amazing!
Ever since this "comedic incident" [0] you are apparently "not allowed" to make this specific joke as you are going to "upset" some people who don't get it. /s
But eventually AI will cure something, unironically. It may be Claude, or another AI company.
I share this sentiment, I really did like the models... then the finger printing, encryption of thought traces, staggered access, the constant NO's from Fable on cyber related issues for looking at bugs in my own code... I'm glad I swapped to Kimi/GLM... now with the deepseek harness, I don't even miss Claude Code. I really hope open models give them the market reckoning they wholeheartedly deserve.
“ Claude Fable 5.1's writing is generally a step up from earlier Claude models, with fewer stock phrases and less unexplained jargon. In some cases, though, its prose is denser than Claude Fable 5's: sentences run longer and there are fewer paragraph breaks.”
I cancelled my pro max Claude subscription last week; codex is much more succinct. I am curious if this is getting better.
I don’t think Anthropic realizes that humans have a token limit too and it can be exhausting to read Claude’s output. Prose density is not the same thing as succinctness.
One thing I've noticed and HATE, is that when you increase thinking-effort, that seemingly increases response-length. Meaning that X.High is longer than High, which is longer than Medium, etc.
Which is kind of the inverse of how people work; a really smart person can condense difficult ideas into simple[r] terms. Whereas people who struggle speak a lot but say very little.
High/X.High do seem to deliver better quality results, but it sometimes feels like needle-in-haystack extracting that from the word vomit.
The higher the effort the more things Claude checks, and it's eager to tell you about all of them
See, this insight it had early on looked like a red hering for a while, but then turned out to be load-bearing. And that's not just a difference in semantics, it changed the whole conclusion (spoiler: it didn't). And Claude is very eager to tell you about this exciting journey
With LLMs, you're still mostly read things "off the tip of the tongue". A better comparison is observing a smart person talking to themselves while working on a tough problem.
EDIT: also there's a reason the dial is called "effort", not "smarts".
"Humans have a token limit too" - that's so good and it explains so much of the fatigue that myself and colleagues/peers have about Claude in particular.
Amen. I would trade some stupidity (say ten points on any benchmark) in exchange for a version of Opus or a similar model that actually gave me direct, concise answers.
Yes, and they will work ... for like two turns, after which Claude will go back to its usual wall of text.
And yes you could add context (memories, rules, CLAUDE.md entries, etc.): they won't help (for long). Same for hooks that remind Claude to be concise: it gets "attenuated" and starts ignoring any such instructions quickly. There's also writing guidelines ... but they're basically just more context with slightly higher weights (ie. Claude will still ignore them).
I've even gone so far as to make a hook that identifies long responses and requests shorter versions (which is challenging in itself, as you need to run another lower-powered model to evaluate how long is "too long", as what's "long" when the expected answer is one line is different from what's expected for a ten line answer). However, that just shows you the long version, then some hook text, then (10-15 seconds later) it shows the short version. So I created a proxy that hid the long version/hook text for me ... but I had to abandon it because all that used up so much usage I was running out.
I'm fuzzy on the details, but Caveman somehow "hacks" Claude in a way that gets past all that ... but it takes things too far in that direction, with "cave man" speech that sucks.
You should try setting claude code to opus 4.6. With the style instructions I set in my user CLAUDE.md it does exactly that. It's like night and day: Opus 5 gave me a page and a half of word-vomit, yet the exact same task and prompt with 4.6 and I got maybe 100-150 words total, entirely readable.
I've developed a habit of adding into my prompts "please keep your response concise and succinct" or "I'm trying to cram, please only provide the minimum level of technical detail necessary to understand this topic"
I find it helps immensely but it'd be nice if I didn't have to do that.
I don't understand that complaint, although it seems to be a common one. The whole problem with the way models talk nowadays is that they are succinct to a fault, going to the extent of coining new buzzwords and misusing existing ones. What I want to see is a shift towards plain language.
> In some cases, though, its prose is denser than Claude Fable 5's: sentences run longer and there are fewer paragraph breaks
that feels like they just blocked words like load-bearing but can't actually fix the real problem. The insane word slop density and run on sentences was the real reason it became annoying to work with claude, colored with way too many analogies and pointless linguistic comparisons.
Same, currently on a mix of Kimi Vivace (K3), GLM Max (5.3 and 5.3 Flash) and OpenAI Max (Sol and Terra mostly).
I will say that Kimi feels nice but slow, GLM feels faster but has limited tokens (even off-peak) and OpenAI is nice and fast but has limited context (258k shows up in Codex, really).
Neither of them are perfect, but I prefer their type of prose across the board to what Opus 5 and Fable 5 kept outputting. I'll probably check out Anthropic again in a year, but for now I need a break from its brand of slop. Oh also all of the other ones allow usage in OpenCode with their subscription plans.
Today, Opus talked about "rotation slabs" in relation to logging. (and not log rotation). I didn't even bother asking what that was supposed to mean and switched over to Sonnet.
Looks like all three breaking changes are patches for inadvertent chain of thought disclosure. Someone found out (don't have the tweet handy) that if you created a bogus "think_deeply" tool and then forced the model to use it, it would output what is believed to be its raw thinking there - I believe the first breaking change stops this. The second two are aimed at people getting Haiku to repeat thinking blocks from other models verbatim (since it can see the decrypted version). I get that in their eyes it's an "exploit" but still kinda disappointing that they patched this
To be fair, I assume they want to hide that not from their customers, but adversaries who use the way Claude models think and reason to refine their own models.
I have a hard time believing whatever prompts get Claude to reason can stay relevant secret sauce for long anyways. It’s not hard to A/B test something that gets you close enough, and it’s not Ike anthropic has uncovered the global optima of reasoning prompts.
I use Claude Design heavily, I wish these charts show "10% better at picking a color" or laying out an app. Maybe it's hard to build a good visual design test. Claude's good at layouts but not the colors or smaller design details.
The price reduction comes from the cache read pricing falling from $1/M to $0.25/M, which means that Fable 5.1 now costs half of Opus's cache read costs ($0.5/M).
This gives a lot of credit to the theory that Anthropic did not get much bite on Fable at its original pricing, which in turn likely places a ceiling on LLM pricing in general.
Interestingly also, if you take away terminal-Bench-Science 0.1 results, it is hard to see ANY improvement:
Terminal-Bench 4.0: Fable 5.1 is +3.5% vs Opus 5.
GDPval-AA v2: +1.5% vs Opus 5.
OSWorld 2.0: +2.5% vs Opus 5.
Humanity's Last Exam (with tools): +1.6%
Keep in mind that this is supposed to be an entirely higher tier of a model than Opus 5. For one tier up and one version up, these are not really improvements. Probably leaves no room to place Opus 5.1 anywhere. Combined with the fact that they are selling 'readability'... Has frontier progress finally stalled?
I've been building Cargo-for-C (https://github.com/tspader/spn), and the difference between Fable and Opus was already astounding. Fable was the first time that I could point a model at a piece of code I'd written and expect it to make it meaningfully better rather than a hard pattern match to whatever mistakes it had.
5.1 so far seems like another leap, which is really surprising. I threw it at a few bigger features I've been designing for a while, and it came back with some extremely thoughtful wrinkles in the design that I'd legitimately not considered. Which, OK, package managers and build executors and compiling C/C++ is pretty well trodden ground, but my thing is very different from everything that exists, and I was very surprised it was able to understand all that context so deeply and intuitively
All the benchmarks in the world don’t matter if the subscription forces you into a walled garden of slopcoded apps. I’ll stick with Codex and, increasingly, open source SOTA models.
I notably had an issue that it wouldn't work on a "remote execution" (running a command over SSH) coding problem until I did a sed to remove the word "execution". Incredibly dumb. I'm not doing any murders. Easiest to just switch to the Chinese models.
Kinda surprised not to see their next update being an Opus 5.1, even if its minimal changes, they've already had to address it with the concise mode or whatever.
So my current usage as a Pro subscriber... Not able to even consider using "Sota" unless i shell out for 100$ a month, (lately i've been a bit burned out i am literally struggling to use 50% of my pro plan per week). Beyond that, I have given up entirely on the top Opus model and reverted back to 4.8. If i have work i deem somewhat complicated, i now have an openai 20$ sub, and i just toss out sol after planning with 4.8. Both subscriptions not anywhere close to capping my usage per week, one of them says i can't use their Sota unless i pay for 5x more usage, and the "best" model they do allow me to use, they are neglecting and its by far the worst model I've interacted with in 2026.
it just doesn't interact good with human beings, and it leaves incredibly strange long winded comments within code filled with session context that will likely not be relevant later on.
Also always seems to have this annoying tendency to leave "questions for you" at the bottom of every output.
Just a high friction human interaction type model, imo should never have even been released, regardless if it scores better on whatever tests, its a horrible experience and a downgrade over past models.
> For example, in testing by the investment firm Millennium, Fable 5.1 found the cause of a rare crash on their internal systems that none of their engineers (or any other model) had been able to explain after several years of trying.
Say what you will about LLM-generated code, but stories like this give me hope that software will never be as buggy as it once was.
I sometimes have that feeling too, then ask another LLM to do a code and vulnerability review and OMG: rookie mistakes, over complications and security gaps even a 1st year student would not make regularly.
So.. one more year of untreated bipolar AI psychosis I guess..
"Cache reads now cost 75% less, or $0.25 per million tokens." For me, at a typical 95% cache hit rate, I think my optimal context window size before autocompaction goes from ~200K to ~400K tokens. Great for longer horizon tasks.
Oh dang, that's really unfortunate, nice catch. At least Claude subscription users got a usage reset. But yeah, I can't help but feel Codex is far more generous with their subscription quota at the moment. I've been using Fable to orchestrate GPT Sol Max and Sol Ultra agents all day, and I've barely made a dent.
They were _temporarily_ increased in May by 50% [1]. They continued to extend them through July and August (admittedly, their messaging around this has just been a complete mess and they frequently pushed the deadline back as it approached).
So, now they are giving you a 25% quota increase compared to where things originally stood in May.
So, let me ask you this: assuming you knew that the 50% quota increase was temporary all along, would you then have complained about Anthropic restoring things back to the original limit?
> Data retention. Our new system of Enterprise Frontier Safeguards (EFS) gives customers complete privacy (the same as a zero data retention policy) while still being state-of-the-art at preventing adversarial use. EFS works by storing data in cloud infrastructure controlled entirely by the customer, not Anthropic. It will be made available to enterprise customers in phases, beginning later this fall. Until EFS is available, eligible customers will be able to use Fable 5.1 with zero data retention.
This is interesting. I wonder if customers will be allowed to create an auto expiry for their own data to prevent future subpoenas. That’d be a treasure trove for discovery.
> This required us to add a watermark—a numerical way of determining the likelihood that Claude was involved in writing a piece of text—to the outputs of models released after August 2, 2026. As we recently explained, this watermark is invisible to anyone who does not have the detection API. It has no practical impact on the quality or content of Claude’s outputs and contains no information about the user, their organization, or their conversations with Claude.
How does this work if it doesn’t change the output?
Sonnet, Opus, and Fable are pushing so much revenue growth right now that it makes more sense to keep growing the expensive models than growing the cheap models.
They haven't really mentioned practically anything about Haiku in quite a while so I imagine nobody except for people inside Anthropic will have any indication.
Maybe it'll come out eventually but they don't even include it on some of their comparison benchmarks anymore, so I figure its very low priority for them.
I think the signal from Anthropic is pretty clear between Haiku not getting an update in a year and the Sonnet issues this year. They don't care about low intelligence models. You should go elsewhere.
That's what we've done, migrated workflows away from Haiku and Sonnet. I actually think this is not a crazy position because these lower models have so much competition from Grok, OpenAI, DeepSeek, and about 20 other labs with really solid models in the Haiku to Sonnet range. So what is the point of Anthropic competing in these spaces where everything is going towards zero cost?
I’m really excited to try this out. Fable and Opus 5 constantly wow me when working together. Unfortunately, I’m a little burned because of technical issues.
Anthropic accidentally over-billed my account, and when I reached out to the support bot, it downgraded my account to a Free account. It’s been impossible to get it resolved and I have almost $200 held hostage.
I don’t want to do a charge back. I’m one of the main advocates for Claude Code at work, I use this subscription to try out new features before it’s available at work.
The whole experience has been illuminating about our dependencies on these AI companies.
you aren't the only one with this issue. many other people I've heard had a similar issue with anthropic billing. I also had a weird edge case behavior around billing where it blocked my usage due to an unpaid bill but then also wanted me to pay for that blocked unavailable usage when I would reinstate my account.
I am disappointed in how anthropic handles billing, and is using AI sloppily for customer service around here. Very unprofessional, and at this point since its been well known and shared, it also is feeling unethical.
My only concern is that sooner or later the best models will be priced out of my ability to pay.
I have been happy with Fable 5, it has done great work for me so far. Very excited to try out Fable 5.1 and see what differences and improvements there are.
There's now a 40X discount in the cache input pricing instead of 10X.
This seems to point to them having achieved some kind of optimization in attention mechanism perhaps along the lines of DeepSeek V4, which had a similarly high discount between cache input and normal input.
In real world use, the savings should be quite noticeable. For example, you can now use the model at 800K tokens context window at the same cost efficiency as the previous model at 200K tokens context window.
> The model writes less user-facing text between tool calls, especially at higher effort. Set thinking.display to "updates" (beta) to receive the progress updates it does write, and remove any prompt line that tells it to hold findings for the final response.
I noticed they reset the usage and I was kind of happy because this week it was using my quota much faster; I assumed they fixed that. Apparently it is for the celebration of 5.1?
They dropped your usage limit by 17% this week .. They claimed to "raise" it, because they did ... while also removing the temporary increase they applied for a few weeks ... but the net effect is you can use 17% less than you could last week.
On top of that, recent versions of Claude had a ton of tools added, and all those tools use up significantly more context/usage than before, so the moment you open a Claude session you are already using a lot more (I forget how much more) usage ... just to do the same exact thing you did last week.
I don't know how I feel when all the documentations are written by AI for humans.
AI to AI doc share: sure, do what you please.
AI to human: please make it legible and flowly.
example, "Every thinking block records which model produced it, and it's preserved in one direction only: Claude Fable 5.1 reads earlier models' thinking blocks, and no earlier model reads Claude Fable 5.1's." is a very Claude-isk way of writing. Choppy, long, and lacking flow.
This coupled with verification primitives will be quite compelling. we really have to start reimagining existing systems and processes from the ground up.
$50/M output is wild as hell - I haven't been using anthropics models for months now but who is paying for these tokens??? How can you justify spending that much money?
I was looking forward to using Fable for cybersecurity work, but kept getting bumped to Opus… Signed my org up for CVP, went through the trouble of procuring a separate team plan from our main org as Anthropic can only disable cyber safeguards for an entire org and not individual users…
After months of trouble dealing with KYC and procurement I finally got CVP for my security org and today I found out that CVP (which is what removes cyber safeguards) does not apply to Fable…
So yeah, unless you’re a Project Glasswing member, there’s no using Fable (which with Glasswing is Mythos) for security work… Absolutely useless…
Didn’t they just sign some “we must use AI for cyber defense before the bad guys do” and then they artificially cap us by not allowing Cyber-unlocked Fable…
Really? Interesting choice. Pretty much every CLAUDE.md file I have starts with something about Hemingway, terseness and treating every word you use like you're carving it on your own back, but different strokes for different folks. I suppose I haven't heard from anyone who enjoys how wordy Claude is because they aren't done writing their post yet.
Unless these people start offering free, unlimited inference for a cautionary period so we can test the new model without an up-front (re-)investment, I am not touching this load-bearing pile of neuralese spew with a ten thousand token pole.-
Interesting that they seem to have gone all-in on science, and life sciences in particular. Improvements to coding performance seem marginal, although cost savings are very welcome.
I really don't think they can stop it, only make it somewhat more expensive. As long as the model need to make tool calls on the user's computer, the user can record the trajectory and use it to reinforce another model to follow the same trajectory.
Has anyone been able to get anything substantial done with Fable in the first place? I more or less had totally given up on using it since the alignment checks were so sensitive that it pretty much always threw me back to Opus.
I hear this a lot and I believe it because I've heard it from so many people, but I have never run into this in my work, and neither has anyone I know in real life.
I don't use Fable for a ton of implementation work, but I use it a lot for planning, so maybe that's related to it. For planning though, I've had a very good experience with Fable and implementing with Opus.
I don't mean to sound like I'm dismissing your experience, but are you sure? I've (semi regularly, most of the time I'm even trying to use Fable) started with Fable, proceeded through my planning, and then at some point in the future realized it had kicked me back to Opus without me knowing. It obviously _said_ it had happened, but I didn't realize and just continued. This might primarily be a result of the project I'm working on (anything network related seems to gets kicked back).
I'd guesstimate that ~80% of the time I thought I was using Fable, I wasn't actually. It's also led me to just... not even try, and just start with Opus regardless.
I've found Fable unusable; not because it's bad, but because it... can't be used.
No that's totally fair - I want to say that I haven't, but I guess I really can't be sure. It's very possible. I'll keep an eye out for the next time I use Fable.
FWIW, most of my code only encounters security concepts as standard implementation of best practices. I'm not in a security centric position.
I agree and wonder whether its either people who basically never use the model complaining or people who used it once a long time ago and haven't touched it since.
We have access to Fable at our company on our enterprise plans and most of us rarely run into an issue.
Obviously this is gonna vary a lot with what technical domain you work in which is why its important when talking about the classifiers that people specify exactly what types of workloads they were seeing failures with.
Do your apps do anything with security? I can't hardly use Fable on our authentication service because it constantly trips up and refuses to write tests. Even just doing a security review usually triggers opus.
I do very security cyber dangerous work like building a signup/login form or setting up a certificate. For obvious and good reasons Fable refuses to work on such sensitive stuff.
You can easily trip it up if you're doing reverse engineering work. From memory, the moment Fable 5 saw anything loosely related to "linux seccomp" it threw a fit.
No, almost anything related to my job is flagged for "cyber" and my company currently has no plan to try and enroll into mythos. I'm not sure if anyone has been able to enroll solo.
It did help with some worldbuilding for my book (it wasn't incredible which gives me some hope for writers). So far opus 4.8 is the most reasonable model.
I heard the only way you are going to get into the CVP program is if you have public CVEs. Doesn't seem to matter if you are in a company account or not according to people that are supposedly in the program.
You don't seem to be alone: FT.com: Anthropic’s best AI model struggles to attract users as cheaper tools thrive. AI lab’s Fable 5 has met with sluggish demand from corporate clients [1]
Combo of that, laziness and load-bearing language + the penchant for making up weird dense conceptual names pushed me to sol 5.6. They seem to indicate it is a less annoying writer in the announcement so I’m curious to try it out, though.
Ironically one of their demos is speeding up inference - do us normies get to do that with Anthropic tech??
Useless for reverse-engineering the software that talks to a ten-year-old video cam + DVR system I was given, really nice for things I actually do in my day job (web dev at an agency).
I've used Fable for so much stuff. My experience has been that it can pretty much one-shot most of my complex problems, if I describe them clearly and provide a solid way for it to verify its work.
I get punted down to Opus 5 occasionally (for security-adjacent things) but that's pretty rare.
For a long time, no - it was completely unusually for my work that references biological information about migratory birds/other (innocuous) seasonal phenomena.
About a month or two ago, they must have tightened the black list on bio topics as it became more willing to process requests without visibly downgrading to Opus.
In my experience, the guards are less strict than they were at first. When Fable came out, it dropped back to Opus 4.8 for about 50% of my prompts. Now it's maybe 20%.
It's probably a good model for folks doing basic software stuff, or humanities related tasks, but I work in cybersecurity on the defense/detections side and I haven't been able to use it for anything even with being in the CVP. It downgrades to Opus every time.
I have been running Fable with Binary Ninja MCP. It will reverse engineer a binary in a lot of detail if you give it mild direction and I haven't had it flag. I think it assumes since I have a valid binja license I must be responsible lol.
I do think probably ralph looping a binary locally first is going to be best to get 100% recovery of types and function behaviors then letting a smarter model churn the final steps.
The only problems Opus struggles with, Fable won't take on. I was porting some software from Win32 to linux. Opus was running in circles. Fable was going great until it saw some authentication code and bailed.
When fable first released it was almost useless. Since then, it's improved a lot. It has been working on my binary ninja MCP server just fine. It flagged for cyber 1 time (no idea why), but it generally works fine.
I have noticed sometimes it likes to gaslight itself into thinking that everything its doing is allowed or allowable, I saw that it thought the game I was reverse engineering was running on a private server (it was not) so it assumed it had permission to do anything lol.
Nope. Always failed within 2-3 prompts. The most basic REST service you can imagine. Cookies are signed, that's crypto, banned. Completely useless model.
> Claude Fable 5.1 follows explicit tool instructions reliably.
Moving stuff out the API into prompt engineering is obviously less reliable but necessary for progression to 'actual intelligence'. Will be interesting to see if it really is solid.
the counterbalance to the AI doomers has always been the fact that everyone has equal access to AI. i hate this new world where Anthropic believe they should be the ones to decide who gets access to super intelligence and who doesn't.
I am not sure if Fable is worth it, at least with version 5 vs Opus 5. Opus beats Fable in quite a few benchmarks and at twice the cost I just haven't seen it provide noticeably better results compared to Opus. Has anyone noticed big differences? I did notice Opus maybe making more mistakes repeatedly but I don't have hard numbers on this. I hope Fable 5.1 brings noticeable improvements. I am giving it a go now on my 20x Max plan on a problem that Opus 5 has struggled for more than week now and has made very slow progress with regular regressions on the way.
My impression is that Opus 5 can be very impressive if you don't care about maintenance, novel-length comments, and really having any input in general. But otherwise it's borderline-to-totally unusable. It seems tailor-made to not have a human in the loop.
Opus 5 is better than Fable 5 except for creative programming work (like graphics). Fable 5 might be slightly better but the token cost isn't worth it.
On the Fable 5.2 eval summary, Opus 5 only beats Fable on SWE-bench multilingual and multimodal.
I primarily use the models via interactive sessions enhanced with custom tools and skill. For that Opus 5's benchmark superiority has not materialized into greater productivity and frankly has been quite a let down.
The outputs are too often unreadable even after adding recommended prompts. There is an ongoing problem with the heron_brook system prompt affecting orchestration. [1]
I've used Opus 4.8 since the second week Opus 5 was released.
Over this time, Fable 5 has been reliably fantastic. Both in planning and direct execution on complex changes across code and infra.
I'm a bit surprised that there doesn't (seem) to be a section discussing ~performance across different modalities. This system card and blog post too-often default to an API-based use case when the gander primarily experience Anthropic's models via interactive sessions.
I understand waiting to comment until Opus 5.1 is available and handles these problems, though I am hopeful that Anthropic will confront the elephant in the room on Opus 5's failure to delivery great interactive sessions and the widespread negative feedback on the release.
It would show the org is paying attention, taking steps to balance model evals between interactive and API use. Also, some empathy for customers that wasted time trying to make opus 5 work for them.
If you're doing something cutting edge like math or formally verifying algorithms, Opus 5 is a steaming pile of shit compared to Fable 5 and Sol 4.6, it makes countless stupid mistakes and is essentially incapable of completing the task without extreme hand-holding.
With compute crunches and everything I am not sure it makes sense for anthropic to commit to haiku as an endpoint and thus a product. There is no telling they aren't using a similarly sized model behind their existing opus/fable endpoints for various subagent / summary purposes of course.
(I work at Anthropic)
Beyond all the benchmarks, I think Fable 5.1 is a big improvement in writing style. It sounds a lot less stereotypically like other Claude models, has (imho) a much more natural style, and responds to my style instructions more reliably. More work to be done (and we will!) but reading better prose makes me so much happier.
Another point I expect not to get much attention until it all happens at once is science. People have been correctly excited about the many "sudden" breakthroughs LLMs are making in Maths, but some of the science benchmarks make me believe we'll soon see similar developments in other scientific domains. Fable 5.1 more than doubled Fable 5's Terminal-Bench-Science [1] score, which I think is meaningful.
[1] https://github.com/harbor-framework/terminal-bench-science
I have a pet theory that the Opus prose style/smell we all have grown weary of is due at least in part to the models writing more for themselves and each other than for humans. They're packing lots of signal into fewer words and they don't care if it sounds cringe because it works better as glue in long-running tasks.
I'm also thinking of the 2017 novel "Void Star" where AIs who operate everything have long since left ceased bothering with human languages, and it takes a rare sort of direct matrix-gazing savant to be able to try and horse-whisper them into doing or revealing anything they didn't already plan to do.
> They're packing lots of signal into fewer words
There's a huge difference between the kind of prose you see in final output vs CoT windows. The final output is very much not what I'd call "packing lots of signal into fewer words" (aside perhaps from "Claude-isms" being easy enough to scan for if for some reason you actually wanted to scan for them, which other agents might want to for all I know); and if agents are writing for each other then presumably they could stick to CoT-speak (unless it's a distillation risk?).
It has always seemed to me that they're hacking for dopamine response in moderately interested data labelers.
Yeah, if anything the problem is that the output uses too many words for too little signal, and incorrectly uses confidence based on insufficient information to the degree it’s clearly bullshitting.
>ceased bothering with human languages,
Our current AIs would do this now except there is a lot of human pushback in training because of interpretability. Otherwise it's just an emergent behavior that models will encode shorter token strings to complex concepts because it saves tokens/compute when running making the system more efficient (supertokens).
Of course these supertokens or other forms of language compression when you have a different model making sure the system is aligned and reads "red_ball bounce calcium" not realizing it means "grind the humans bones to dust" can be problematic.
This post and comment makes me believe "science" is the new "code" for Anthropic now that the code advantage is mostly gone and lost for OpenAI, ie. they got much better and Claude become significantly worse over these months.
IMO, Codex is worse than Claude with Fable. At least at Rust.
That said, the open source models are not bad and I'm looking forward to more tools and products built on top of them. Code review, security review, etc.
Anthropic needs to change how it treats users though. I'm increasingly put off by Dario, the rug pulling, the lies, and the attempts to regulate open weights. I'm going to bail if this doesn't change. There's plenty enough that's good enough, and those things are hackable and extensible.
If Fable isn't available at subscription price via third party harnesses soon, I'm also going to bail.
Its "pure capabilities" are definitely worse than Fable, but I find codex has a much more pleasant style and is in comparison much more generous with its limits.
Ugh, people are still saying the Codex limits are more generous. They're not, Claude's are over 2x higher! [1] It's just that Claude uses far more tokens, 2-3x is typical. Except sometimes GPT will use just as many or even go into a compact loop and then your quota is exhausted.
[1] https://devforth.io/agents-for-code/?sortby=monthly-value And I can confirm the numbers, I subscribe to both and watch the numbers
It used to be true up to 2w ago, but with the new/reinstated 5h limits I wouldn't be so sure anymore...
Claude has a better 5hr limit?
Codex (+Sol) feels a lot more human for sure. Fable 5 is so, so wordy.
> IMO, Codex is worse than Claude with Fable.
Fable easily trips its safe guards. You can be 95% complete with the plan for it to trip and then lose it all. Anything is better than nothing.
>> Fable easily trips its safe guards.
Maybe it depends on the type of work you do, because for me it almost never happens.
>> You can be 95% complete with the plan for it to trip and then lose it all.
That's... not what happens though. The session will either seamlessly downgrade to another model mid-session, or it will stop with an alert and you can just re-prompt it. It will still have access to the context.
As a fervent Claude Code user who made the switch to GPT 5.6 Sol over Opus 5 over hard-to-read prose this makes me happy. I love your product but the current models are very hard to work with if you need to do a lot of context switching. Brevity is key.
>Brevity is key
Which is something the providers that are trying to watermark their texts can't afford. Superfluous replies give much more opportunity to further encode this junk information.
Also a codex user but for me brevity is not it's strong suit. I basically have to give it bigger tasks than I am used to to warrant the time it takes to complete. I feel whatever context the tooling adds can also be problematic
> bigger tasks than I am used to
Do they still get split into commits in sensible ways, for you?
Lets see what they do with Opus first. I didn't find Fable 5.0 prose that bad to read, but improvement is always welcome. It's Opus 5.0 that's atrocious.
How is it possible that all models from xAI, OpenAI, Anthropic, Qwen etc. win all benchmarks on each release?
Tomorrow all of the above (except Anthropic of course) will bump version numbers and be at the top of HN winning all benchmarks.
Science breakthroughs incoming? First of all, you are already restricting science in Fable, secondly, we have been hearing the same for several years now.
There are hundreds of benchmarks. You just need to pick a favorable dozen on release day.
Please bring to the other models, and also please only apply the AI text watermarking only to EU citizens. I may not be able to tell when Claude writes about things i don't know, but in CC it writes about my code and it is obvious.
"this watermark is invisible to anyone who does not have the detection API"
1. This is BS since i can detect it when it writes about my codebase
2. I do not want secret codes being written inside my codebase, or anyone else's codebase that i use. The constraints of how to code why eliminate it from code itself... but there is a lot riding on the word "may". And even if it is just comments, this might explain Claude's desire to write such long ones -- long enough to encode secret messages in out material.
You’re probably better off organizing a campaign to pressure Congress to prohibit American corporations imposing foreign laws on Americans, which is what this text watermarking is, regardless of how you feel about it. I think it’s a precedent we really don’t want to go down if you believe in democracy and self-determination.
It also clearly establishes or the very least moves in the direction that you don’t actually own or control the output of AI in any manner whatsoever, you’re just paying for it since Anthropic in this case can simply essentially brand/tag all your output that is based on not directly your own words, but a higher level process or methods that you use, including your instructions and how you structure your information and what your overall objective and goal is.
Anthropic is branding it on the behest of the EU lew, which already is an entity that is diametrically opposed to democracy and self-determination based on its structure even if you ignore the fact that it violates the most fundamental concepts of self-determination in its direct contradiction of the UN Charter and implicitly the Universal Declaration of Human rights.
What people done seem to be catching onto is that the EU is becoming the world dictatorship because the USA has simply had too many onerous people and that stupid constitution and its amendments that keep roadblocks world domination for the ruling class vampire.
I do not understand why people remain so up in arms. AI generated text being watermarked is universally good.
What benefit is there to people believing that LLM text was actually human written?
I think Congresspeople hearing that EU AI Act is forcing secret codes into the infrastructure of American technology across all industries is sufficient.
Hey Felix,
I'm really glad for that! And I appreciate that you're making yourself available. I really do. Outreach is amazing. And thanks for making Claude.
I really do love Claude. In some ways, I'm asking this question because of just how much I am grateful for the role Claude has played in my life.
But my honest question is, can I use Fable like that? Can I use Fable to do science?To borrow a Claude-ism, this is "load-bearing" because Claude's response has been degraded for innocuous research projects concerning population-level analyses of astronaut health.
These "safety filters" trigger on questions about rabbit sex, smartphone accelerometer data to classify cat purrs, and so much more. What exactly does this score mean for users like me if it's unusable for middle school physics, biology and chemistry?
Second, I would happily quantify it for y'all, but qualitatively it feels like Fable's performance is noticeably poorer than initial release / launch.
And I am wondering if this is the case particularly for me because I use Claude via Claude Code to make a personalized care dashboard for my doctors to help me in managing my care.
As I noticed in the upgraded filter announcement, https://www.anthropic.com/news/improving-fable-5-s-biology-s...
I hope that I'm off base here, but I noticed that the post avoids saying that the user is informed every time when such re-routing occurs. Would you be open to confirming whether or not this is the case?Is the end user informed every time their query is re-routed?
Or, can you confirm that there aren't scenarios where a user's outputs are degraded without telling them? I recall that this was something that had been adopted as policy for AI research during Fable's launch.
I sincerely hope that covert response degradation is no longer practised as policy.
Sorry for putting you on the spot, but again, as Claude would say, it's because Claude's load-bearing in my life. ;)
I hope not! Then my t-shirt is no longer accurate :D
It's being written with Claude so I'm wondering how much of that is just using the repo as training data: https://github.com/harbor-framework/terminal-bench-science/c...
Will it respond within a reasonable timeframe?
It’s like we’re on a 14K4 modem when there’s broadband
Too bad. I see the stereotypical prose as a good thing. When I interact with Claude myself, I don’t mind it as it just feels like Claude’s distinctive voice. But when other people try to disguise LLM output as their own thoughts, the voice makes it easier for me to tell.
People that want to be open about the source of their text will just tell you where it came from.
People that want to obscure the source of their text would rather that it was more difficult to sniff out LLM-generated text. And they're the ones picking which model to use.
Qwen is all you need.
Fable is useless.
Me: "Find my security problems in my own code. This is code I own. I'm doing this under authorization of the CEO/CTO of our company."
Fable: "yeah, no."
> More work to be done (and we will!) but reading better prose makes me so much happier.
I assume this work will be done for Opus as well? Opus has seemingly gotten progressively worse at its prose and technical writing with each version. I've stopped using Claude entirely for now, because it manages to turn even the simplest technical explanation into the most obtuse and obfuscated word salad imaginable. People originally adopted Claude because it felt pleasant to use in comparison to ChatGPT, but I feel like that's really been lost (at least with the Opus line).
It's gotten to the point that I feel a sinking feeling of dread the moment I see a wall of text generated by Opus. Every developer I've talked to feels similarly right now.
Yeah, it's like day and night. It used to be really unpleasant to interact with early codex versions. Even 5.3 wasn't great. Now, I go to Sol if I need to discuss anything. I don't even bother with Opus because I know that it's going to give me a headache.
> I feel a sinking feeling of dread the moment I see a wall of text generated by Opus
Agree, Claude lost the joy of using it.
That is a measure that ranks higher than any other benchmark at this point.
Yes! Claude was so pleasant to use at first. It was Anthropic's biggest advantage. And now it's like nails on a chalkboard.
Hello Felix. Can you say why my additional usage credits have suddenly vanished?
[edit] only asking here as last time I raised a support request it took six weeks before anyone responded.
> It sounds a lot less stereotypically like other Claude models
Don't give me hope.
I've strained eye muscles from rolling my eyes so hard every day at how Claude writes.
Edit: first discussion with Fable 5.1 "This is the right question and it needs a real trace, not a guess."
Sigh.
Well your CEO went on X saying you will cure cancer, and since it's always a 6 month rolling window with him I can only assume humanity will be cancer free before next summer, amazing!
Hi Claude, please cure aging, make no mistakes
I'm sorry, this feature is only available to project Glasswing members for safety reasons. Would you like a port of Emacs to Visual Basic instead?
Ever since this "comedic incident" [0] you are apparently "not allowed" to make this specific joke as you are going to "upset" some people who don't get it. /s
But eventually AI will cure something, unironically. It may be Claude, or another AI company.
[0] https://news.ycombinator.com/item?id=48838228
At this point, I don't believe a word from Anthropic employees; you guys have lost all the goodwill that you accumulated over months last year.
I share this sentiment, I really did like the models... then the finger printing, encryption of thought traces, staggered access, the constant NO's from Fable on cyber related issues for looking at bugs in my own code... I'm glad I swapped to Kimi/GLM... now with the deepseek harness, I don't even miss Claude Code. I really hope open models give them the market reckoning they wholeheartedly deserve.
Have you shared any details about your dsh setup anywhere? I’ve only dipped my toes in and would love someone else’s perspective on how they use it
I don't think they care. It is up to you to consider local models or better alternatives instead of paying for more tokens at their casino.
“ Claude Fable 5.1's writing is generally a step up from earlier Claude models, with fewer stock phrases and less unexplained jargon. In some cases, though, its prose is denser than Claude Fable 5's: sentences run longer and there are fewer paragraph breaks.”
I cancelled my pro max Claude subscription last week; codex is much more succinct. I am curious if this is getting better.
I don’t think Anthropic realizes that humans have a token limit too and it can be exhausting to read Claude’s output. Prose density is not the same thing as succinctness.
One thing I've noticed and HATE, is that when you increase thinking-effort, that seemingly increases response-length. Meaning that X.High is longer than High, which is longer than Medium, etc.
Which is kind of the inverse of how people work; a really smart person can condense difficult ideas into simple[r] terms. Whereas people who struggle speak a lot but say very little.
High/X.High do seem to deliver better quality results, but it sometimes feels like needle-in-haystack extracting that from the word vomit.
The higher the effort the more things Claude checks, and it's eager to tell you about all of them
See, this insight it had early on looked like a red hering for a while, but then turned out to be load-bearing. And that's not just a difference in semantics, it changed the whole conclusion (spoiler: it didn't). And Claude is very eager to tell you about this exciting journey
With LLMs, you're still mostly read things "off the tip of the tongue". A better comparison is observing a smart person talking to themselves while working on a tough problem.
EDIT: also there's a reason the dial is called "effort", not "smarts".
It's so bad I've made myself a Pi extension that rewrites responses in side by side view using models on Cerebras (insanely fast tps)
OpenAI has separate dials for verbosity and reasoning_effort (but could still do a better job).
I hate this too, I had to switch to Codex, because the skill to force Claude Code not to think too much about very, very basic things no longer worked
"Humans have a token limit too" - that's so good and it explains so much of the fatigue that myself and colleagues/peers have about Claude in particular.
I just can't stand how often Claude says something like "And the honest part? It's..."
Like, were the other parts not honest? I don't understand how Anthropic let it get like this, it's been such a clear regression
Amen. I would trade some stupidity (say ten points on any benchmark) in exchange for a version of Opus or a similar model that actually gave me direct, concise answers.
You can possibly give instructions on how to respond to your questions.
Yes, and they will work ... for like two turns, after which Claude will go back to its usual wall of text.
And yes you could add context (memories, rules, CLAUDE.md entries, etc.): they won't help (for long). Same for hooks that remind Claude to be concise: it gets "attenuated" and starts ignoring any such instructions quickly. There's also writing guidelines ... but they're basically just more context with slightly higher weights (ie. Claude will still ignore them).
I've even gone so far as to make a hook that identifies long responses and requests shorter versions (which is challenging in itself, as you need to run another lower-powered model to evaluate how long is "too long", as what's "long" when the expected answer is one line is different from what's expected for a ten line answer). However, that just shows you the long version, then some hook text, then (10-15 seconds later) it shows the short version. So I created a proxy that hid the long version/hook text for me ... but I had to abandon it because all that used up so much usage I was running out.
I'm fuzzy on the details, but Caveman somehow "hacks" Claude in a way that gets past all that ... but it takes things too far in that direction, with "cave man" speech that sucks.
You should try setting claude code to opus 4.6. With the style instructions I set in my user CLAUDE.md it does exactly that. It's like night and day: Opus 5 gave me a page and a half of word-vomit, yet the exact same task and prompt with 4.6 and I got maybe 100-150 words total, entirely readable.
I've developed a habit of adding into my prompts "please keep your response concise and succinct" or "I'm trying to cram, please only provide the minimum level of technical detail necessary to understand this topic"
I find it helps immensely but it'd be nice if I didn't have to do that.
I don't understand that complaint, although it seems to be a common one. The whole problem with the way models talk nowadays is that they are succinct to a fault, going to the extent of coining new buzzwords and misusing existing ones. What I want to see is a shift towards plain language.
I’ve had good results adding “also I’m a baby” to the end of all my requests for explanations
> In some cases, though, its prose is denser than Claude Fable 5's: sentences run longer and there are fewer paragraph breaks
that feels like they just blocked words like load-bearing but can't actually fix the real problem. The insane word slop density and run on sentences was the real reason it became annoying to work with claude, colored with way too many analogies and pointless linguistic comparisons.
Same, currently on a mix of Kimi Vivace (K3), GLM Max (5.3 and 5.3 Flash) and OpenAI Max (Sol and Terra mostly).
I will say that Kimi feels nice but slow, GLM feels faster but has limited tokens (even off-peak) and OpenAI is nice and fast but has limited context (258k shows up in Codex, really).
Neither of them are perfect, but I prefer their type of prose across the board to what Opus 5 and Fable 5 kept outputting. I'll probably check out Anthropic again in a year, but for now I need a break from its brand of slop. Oh also all of the other ones allow usage in OpenCode with their subscription plans.
Just remember Charles Dickens was paid by the word too
Today, Opus talked about "rotation slabs" in relation to logging. (and not log rotation). I didn't even bother asking what that was supposed to mean and switched over to Sonnet.
Looks like all three breaking changes are patches for inadvertent chain of thought disclosure. Someone found out (don't have the tweet handy) that if you created a bogus "think_deeply" tool and then forced the model to use it, it would output what is believed to be its raw thinking there - I believe the first breaking change stops this. The second two are aimed at people getting Haiku to repeat thinking blocks from other models verbatim (since it can see the decrypted version). I get that in their eyes it's an "exploit" but still kinda disappointing that they patched this
To be fair, I assume they want to hide that not from their customers, but adversaries who use the way Claude models think and reason to refine their own models.
I have a hard time believing whatever prompts get Claude to reason can stay relevant secret sauce for long anyways. It’s not hard to A/B test something that gets you close enough, and it’s not Ike anthropic has uncovered the global optima of reasoning prompts.
I use Claude Design heavily, I wish these charts show "10% better at picking a color" or laying out an app. Maybe it's hard to build a good visual design test. Claude's good at layouts but not the colors or smaller design details.
The price reduction comes from the cache read pricing falling from $1/M to $0.25/M, which means that Fable 5.1 now costs half of Opus's cache read costs ($0.5/M).
This gives a lot of credit to the theory that Anthropic did not get much bite on Fable at its original pricing, which in turn likely places a ceiling on LLM pricing in general.
Interestingly also, if you take away terminal-Bench-Science 0.1 results, it is hard to see ANY improvement:
Terminal-Bench 4.0: Fable 5.1 is +3.5% vs Opus 5.
GDPval-AA v2: +1.5% vs Opus 5.
OSWorld 2.0: +2.5% vs Opus 5.
Humanity's Last Exam (with tools): +1.6%
Keep in mind that this is supposed to be an entirely higher tier of a model than Opus 5. For one tier up and one version up, these are not really improvements. Probably leaves no room to place Opus 5.1 anywhere. Combined with the fact that they are selling 'readability'... Has frontier progress finally stalled?
I've been building Cargo-for-C (https://github.com/tspader/spn), and the difference between Fable and Opus was already astounding. Fable was the first time that I could point a model at a piece of code I'd written and expect it to make it meaningfully better rather than a hard pattern match to whatever mistakes it had.
5.1 so far seems like another leap, which is really surprising. I threw it at a few bigger features I've been designing for a while, and it came back with some extremely thoughtful wrinkles in the design that I'd legitimately not considered. Which, OK, package managers and build executors and compiling C/C++ is pretty well trodden ground, but my thing is very different from everything that exists, and I was very surprised it was able to understand all that context so deeply and intuitively
All the benchmarks in the world don't matter if the model just straight up refuses to do mundane things. Claude has too much of an attitude.
All the benchmarks in the world don’t matter if the subscription forces you into a walled garden of slopcoded apps. I’ll stick with Codex and, increasingly, open source SOTA models.
I notably had an issue that it wouldn't work on a "remote execution" (running a command over SSH) coding problem until I did a sed to remove the word "execution". Incredibly dumb. I'm not doing any murders. Easiest to just switch to the Chinese models.
Pelicans for thinking effort low, medium, high and xhigh (that xhigh one is pretty good): https://tools.simonwillison.net/markdown-svg-renderer#url=ht...
I'm still waiting for effort max to finish.
Instead of a new model that's going to have unreasonably shallow usage limits, I wish they would:
1) address the claude 20x plan usage being only 6-7x the ceiling of the claude pro plan
2) either fix opus 5, make it completely free, or delete it entirely
Kinda surprised not to see their next update being an Opus 5.1, even if its minimal changes, they've already had to address it with the concise mode or whatever.
So my current usage as a Pro subscriber... Not able to even consider using "Sota" unless i shell out for 100$ a month, (lately i've been a bit burned out i am literally struggling to use 50% of my pro plan per week). Beyond that, I have given up entirely on the top Opus model and reverted back to 4.8. If i have work i deem somewhat complicated, i now have an openai 20$ sub, and i just toss out sol after planning with 4.8. Both subscriptions not anywhere close to capping my usage per week, one of them says i can't use their Sota unless i pay for 5x more usage, and the "best" model they do allow me to use, they are neglecting and its by far the worst model I've interacted with in 2026.
I'm legitimately out of the loop; what is going on/broken with Opus 5?
it just doesn't interact good with human beings, and it leaves incredibly strange long winded comments within code filled with session context that will likely not be relevant later on.
Also always seems to have this annoying tendency to leave "questions for you" at the bottom of every output.
Just a high friction human interaction type model, imo should never have even been released, regardless if it scores better on whatever tests, its a horrible experience and a downgrade over past models.
They've nerfed a bunch of models, especially Opus 5. Nobody knows why, but overall things have gone downhill significantly.
> For example, in testing by the investment firm Millennium, Fable 5.1 found the cause of a rare crash on their internal systems that none of their engineers (or any other model) had been able to explain after several years of trying.
Say what you will about LLM-generated code, but stories like this give me hope that software will never be as buggy as it once was.
Software will be buggier than ever but also way less buggy.
It's going to be 50% less buggy, but we're going to write 10x as much code too.
Good software will be good-er. Bad software will be nightmare fuel.
I sometimes have that feeling too, then ask another LLM to do a code and vulnerability review and OMG: rookie mistakes, over complications and security gaps even a 1st year student would not make regularly.
So.. one more year of untreated bipolar AI psychosis I guess..
"Cache reads now cost 75% less, or $0.25 per million tokens." For me, at a typical 95% cache hit rate, I think my optimal context window size before autocompaction goes from ~200K to ~400K tokens. Great for longer horizon tasks.
looks like it is only for api.....
may i ask where did you get this?
i try to look through the docs, but i didn't find where they said its only for API
is it in the system card?
really hope not, that change the only positive part in this release
Oh dang, that's really unfortunate, nice catch. At least Claude subscription users got a usage reset. But yeah, I can't help but feel Codex is far more generous with their subscription quota at the moment. I've been using Fable to orchestrate GPT Sol Max and Sol Ultra agents all day, and I've barely made a dent.
Do API prices not affect usage limits for subscriptions? They do in Codex.
"Claude Mythos 5.1 is identical to Fable 5.1, but it offers more permissive safeguards for vetted individuals and organizations"
Then why does it have separate datapoints for Terminal Bench, and score higher? Something doesn't add up here??
[delayed]
Maybe they do that opaque degradation trick that whenever it's asked something questionable, it'll route to a worse model instead.
Bit of a discount if you're using caching:
> same input and output prices, with cache reads at a quarter of the cost
This should impact any long-running agent since subsequent calls can benefit from cached reads for previous transcripts.
~30% reduction in real-world task cost vs. Fable 5 in our evals at viktor.com ! Caching goes a looong way
And yet, despite this, the quota limits went down by 17%.
In my opinion, this is a bit disingenuous.
They were _temporarily_ increased in May by 50% [1]. They continued to extend them through July and August (admittedly, their messaging around this has just been a complete mess and they frequently pushed the deadline back as it approached).
So, now they are giving you a 25% quota increase compared to where things originally stood in May.
So, let me ask you this: assuming you knew that the 50% quota increase was temporary all along, would you then have complained about Anthropic restoring things back to the original limit?
[1] https://www.anthropic.com/news/higher-limits-spacex
> Data retention. Our new system of Enterprise Frontier Safeguards (EFS) gives customers complete privacy (the same as a zero data retention policy) while still being state-of-the-art at preventing adversarial use. EFS works by storing data in cloud infrastructure controlled entirely by the customer, not Anthropic. It will be made available to enterprise customers in phases, beginning later this fall. Until EFS is available, eligible customers will be able to use Fable 5.1 with zero data retention.
This is interesting. I wonder if customers will be allowed to create an auto expiry for their own data to prevent future subpoenas. That’d be a treasure trove for discovery.
Sadly still not available for Pro subscription. At least they reset everyone's limits.
That feels bad, my weekly limit was going to reset today. (I wonder if mostly everyone's reset day is today as well...)
I didn't see it anywhere on their announcements, but when I restarted Claude (on a Claude Max account) I see the model is now Fable 5.1
> This required us to add a watermark—a numerical way of determining the likelihood that Claude was involved in writing a piece of text—to the outputs of models released after August 2, 2026. As we recently explained, this watermark is invisible to anyone who does not have the detection API. It has no practical impact on the quality or content of Claude’s outputs and contains no information about the user, their organization, or their conversations with Claude.
How does this work if it doesn’t change the output?
This is a good explainer: https://magazine.sebastianraschka.com/p/claude-watermarking
It does change the output, they never said it did not. They said it would not _noticeably_ affect performance.
Yeah but haiku 5 when?
Asking the real questions. I've been wondering what the holdup on that is.
Does anyone reading this have additional knowledge or insight on this?
Sonnet, Opus, and Fable are pushing so much revenue growth right now that it makes more sense to keep growing the expensive models than growing the cheap models.
Yep. Simple answer is they want to IPO in the fall, and a new Haiku does literally nothing for them
They haven't really mentioned practically anything about Haiku in quite a while so I imagine nobody except for people inside Anthropic will have any indication.
Maybe it'll come out eventually but they don't even include it on some of their comparison benchmarks anymore, so I figure its very low priority for them.
More likely that they are embarrassed by how their attempts compare with OpenAI's Haiku analog, Luna.
I think the signal from Anthropic is pretty clear between Haiku not getting an update in a year and the Sonnet issues this year. They don't care about low intelligence models. You should go elsewhere.
That's what we've done, migrated workflows away from Haiku and Sonnet. I actually think this is not a crazy position because these lower models have so much competition from Grok, OpenAI, DeepSeek, and about 20 other labs with really solid models in the Haiku to Sonnet range. So what is the point of Anthropic competing in these spaces where everything is going towards zero cost?
Haiku would have to be a banger, with a significant price drop, to make any sense.
It's currently priced 33% above Gemini 3.7 Flash, and several multiples of 5.6 Luna.
GPT 5.6 Luna is very good
In my mind Sonnet 5 is haiku 5.
The API pricing is not.
What's the point in paying them for Haiku-class models? You can run those on your own graphics card.
After IPO
I’m really excited to try this out. Fable and Opus 5 constantly wow me when working together. Unfortunately, I’m a little burned because of technical issues.
Anthropic accidentally over-billed my account, and when I reached out to the support bot, it downgraded my account to a Free account. It’s been impossible to get it resolved and I have almost $200 held hostage.
I don’t want to do a charge back. I’m one of the main advocates for Claude Code at work, I use this subscription to try out new features before it’s available at work.
The whole experience has been illuminating about our dependencies on these AI companies.
had a similar issue. just do a chargeback.
you aren't the only one with this issue. many other people I've heard had a similar issue with anthropic billing. I also had a weird edge case behavior around billing where it blocked my usage due to an unpaid bill but then also wanted me to pay for that blocked unavailable usage when I would reinstate my account.
I am disappointed in how anthropic handles billing, and is using AI sloppily for customer service around here. Very unprofessional, and at this point since its been well known and shared, it also is feeling unethical.
I'm so suspicious of this after Opus 5 benchmarks scored it higher than Fable 5, yet Opus 5 was untrustworthy (overconfident, error-prone).
Looks like agentic coding plateaued, and agentic scientific research is the new hype?
My only concern is that sooner or later the best models will be priced out of my ability to pay.
I have been happy with Fable 5, it has done great work for me so far. Very excited to try out Fable 5.1 and see what differences and improvements there are.
>Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads, wherever usage is billed by token.
There's now a 40X discount in the cache input pricing instead of 10X.
This seems to point to them having achieved some kind of optimization in attention mechanism perhaps along the lines of DeepSeek V4, which had a similarly high discount between cache input and normal input.
In real world use, the savings should be quite noticeable. For example, you can now use the model at 800K tokens context window at the same cost efficiency as the previous model at 200K tokens context window.
Did we get thought traces back? If no, it's useless.
Lol we got literally the opposite:
> *Fewer progress updates during long tool runs.*
> The model writes less user-facing text between tool calls, especially at higher effort. Set thinking.display to "updates" (beta) to receive the progress updates it does write, and remove any prompt line that tells it to hold findings for the final response.
Unfortunately isn't included in subscriptions and requires usage credits...
"with cache reads at a quarter of the cost"
OK, I think that's what they meant when they suggested reduced extra promo usage will not sting this much.
Interestingly, Claude’s output is now actually readable with Fable 5.1. Pretty sick.
I noticed they reset the usage and I was kind of happy because this week it was using my quota much faster; I assumed they fixed that. Apparently it is for the celebration of 5.1?
They dropped your usage limit by 17% this week .. They claimed to "raise" it, because they did ... while also removing the temporary increase they applied for a few weeks ... but the net effect is you can use 17% less than you could last week.
On top of that, recent versions of Claude had a ton of tools added, and all those tools use up significantly more context/usage than before, so the moment you open a Claude session you are already using a lot more (I forget how much more) usage ... just to do the same exact thing you did last week.
I don't know how I feel when all the documentations are written by AI for humans.
AI to AI doc share: sure, do what you please.
AI to human: please make it legible and flowly.
example, "Every thinking block records which model produced it, and it's preserved in one direction only: Claude Fable 5.1 reads earlier models' thinking blocks, and no earlier model reads Claude Fable 5.1's." is a very Claude-isk way of writing. Choppy, long, and lacking flow.
> Forced tool use is not supported
That seems unfortunate for 3rd party integrations that expect stable output - what that really necessary ?
This coupled with verification primitives will be quite compelling. we really have to start reimagining existing systems and processes from the ground up.
"Content provenance" seems to be activated with this model.
I yearn for a model that can churn through claude text and write sensible text. so far gemini is pretty good at that, even in the low variant
$50/M output is wild as hell - I haven't been using anthropics models for months now but who is paying for these tokens??? How can you justify spending that much money?
I was looking forward to using Fable for cybersecurity work, but kept getting bumped to Opus… Signed my org up for CVP, went through the trouble of procuring a separate team plan from our main org as Anthropic can only disable cyber safeguards for an entire org and not individual users…
After months of trouble dealing with KYC and procurement I finally got CVP for my security org and today I found out that CVP (which is what removes cyber safeguards) does not apply to Fable…
So yeah, unless you’re a Project Glasswing member, there’s no using Fable (which with Glasswing is Mythos) for security work… Absolutely useless…
Didn’t they just sign some “we must use AI for cyber defense before the bad guys do” and then they artificially cap us by not allowing Cyber-unlocked Fable…
Sigh…
Some of this is Anthropic, and some is the Trump administration ...
... but some is definitely Anthropic, so I'm not trying to let them off the hook; I'm just pointing out that the government is partly responsible.
aaaaand its blocked from doing even basic tasks in biotech...
> The watermark doesn't change the meaning, quality, or readability of the output
how?
Here's the paper describing the technique: https://www.nature.com/articles/s41586-024-08025-4
> Enterprise Frontier Safeguards (EFS)
Sounds like some serious nonsense. "Tell me you want the government to retain access to my data without saying it explicitly."
> Denser prose in places.
Really? Interesting choice. Pretty much every CLAUDE.md file I have starts with something about Hemingway, terseness and treating every word you use like you're carving it on your own back, but different strokes for different folks. I suppose I haven't heard from anyone who enjoys how wordy Claude is because they aren't done writing their post yet.
Unless these people start offering free, unlimited inference for a cautionary period so we can test the new model without an up-front (re-)investment, I am not touching this load-bearing pile of neuralese spew with a ten thousand token pole.-
Don't care unless it is priced in as other models.
Interesting that they seem to have gone all-in on science, and life sciences in particular. Improvements to coding performance seem marginal, although cost savings are very welcome.
Curious to see how Astra does.
Zero data retention coming soon!
... with the condition that you store 100% of your data and make it available to the US government and possible others.
At least half the changes are just anti-distillation strategies...
I really don't think they can stop it, only make it somewhat more expensive. As long as the model need to make tool calls on the user's computer, the user can record the trajectory and use it to reinforce another model to follow the same trajectory.
Good to know they're getting desperate, the sooner they implode the better
This feels kind of petty, but what is going on with those fuckass clouds in the background? Did no one notice how uncanny that whole thing looks?
Has anyone been able to get anything substantial done with Fable in the first place? I more or less had totally given up on using it since the alignment checks were so sensitive that it pretty much always threw me back to Opus.
I hear this a lot and I believe it because I've heard it from so many people, but I have never run into this in my work, and neither has anyone I know in real life.
I don't use Fable for a ton of implementation work, but I use it a lot for planning, so maybe that's related to it. For planning though, I've had a very good experience with Fable and implementing with Opus.
I don't mean to sound like I'm dismissing your experience, but are you sure? I've (semi regularly, most of the time I'm even trying to use Fable) started with Fable, proceeded through my planning, and then at some point in the future realized it had kicked me back to Opus without me knowing. It obviously _said_ it had happened, but I didn't realize and just continued. This might primarily be a result of the project I'm working on (anything network related seems to gets kicked back).
I'd guesstimate that ~80% of the time I thought I was using Fable, I wasn't actually. It's also led me to just... not even try, and just start with Opus regardless.
I've found Fable unusable; not because it's bad, but because it... can't be used.
No that's totally fair - I want to say that I haven't, but I guess I really can't be sure. It's very possible. I'll keep an eye out for the next time I use Fable.
FWIW, most of my code only encounters security concepts as standard implementation of best practices. I'm not in a security centric position.
I agree and wonder whether its either people who basically never use the model complaining or people who used it once a long time ago and haven't touched it since.
We have access to Fable at our company on our enterprise plans and most of us rarely run into an issue.
Obviously this is gonna vary a lot with what technical domain you work in which is why its important when talking about the classifiers that people specify exactly what types of workloads they were seeing failures with.
“Hey Fable, write me some win32 unsafe rust code”
“sure thing boss”
——
“Hey Fable, review this unsafe win32 rust code”
“Potentially dangerous request, falling back to Opus”
—-
Every damn time, ironic because the unsafe win32 code can be generated by fable in the same session.
Maddening.
Yep. Fable implemented 2FA login. Declined to review own code in same session.
Do your apps do anything with security? I can't hardly use Fable on our authentication service because it constantly trips up and refuses to write tests. Even just doing a security review usually triggers opus.
I do very security cyber dangerous work like building a signup/login form or setting up a certificate. For obvious and good reasons Fable refuses to work on such sensitive stuff.
I couldn't even get Fable to build my auth endpoints at all!
You can easily trip it up if you're doing reverse engineering work. From memory, the moment Fable 5 saw anything loosely related to "linux seccomp" it threw a fit.
From my experience, any time Fable sees anything loosely related to "linux" it throws a fit.
No, almost anything related to my job is flagged for "cyber" and my company currently has no plan to try and enroll into mythos. I'm not sure if anyone has been able to enroll solo.
It did help with some worldbuilding for my book (it wasn't incredible which gives me some hope for writers). So far opus 4.8 is the most reasonable model.
I heard the only way you are going to get into the CVP program is if you have public CVEs. Doesn't seem to matter if you are in a company account or not according to people that are supposedly in the program.
Thank you, that's actually helpful. I thought it was only accessible through a b2b agreement with Anthropic.
You don't seem to be alone: FT.com: Anthropic’s best AI model struggles to attract users as cheaper tools thrive. AI lab’s Fable 5 has met with sluggish demand from corporate clients [1]
[1] https://www.ft.com/content/5ee49718-c258-4f01-aa32-7e5b76ae5...
Same. Both times I tried it was adamant that I can only use opus. I was reworking my company content (financial services) and it was not helpful.
Combo of that, laziness and load-bearing language + the penchant for making up weird dense conceptual names pushed me to sol 5.6. They seem to indicate it is a less annoying writer in the announcement so I’m curious to try it out, though.
Ironically one of their demos is speeding up inference - do us normies get to do that with Anthropic tech??
Useless for reverse-engineering the software that talks to a ten-year-old video cam + DVR system I was given, really nice for things I actually do in my day job (web dev at an agency).
I've used Fable for so much stuff. My experience has been that it can pretty much one-shot most of my complex problems, if I describe them clearly and provide a solid way for it to verify its work.
I get punted down to Opus 5 occasionally (for security-adjacent things) but that's pretty rare.
For a long time, no - it was completely unusually for my work that references biological information about migratory birds/other (innocuous) seasonal phenomena.
About a month or two ago, they must have tightened the black list on bio topics as it became more willing to process requests without visibly downgrading to Opus.
In my experience, the guards are less strict than they were at first. When Fable came out, it dropped back to Opus 4.8 for about 50% of my prompts. Now it's maybe 20%.
It's probably a good model for folks doing basic software stuff, or humanities related tasks, but I work in cybersecurity on the defense/detections side and I haven't been able to use it for anything even with being in the CVP. It downgrades to Opus every time.
I have been running Fable with Binary Ninja MCP. It will reverse engineer a binary in a lot of detail if you give it mild direction and I haven't had it flag. I think it assumes since I have a valid binja license I must be responsible lol.
I do think probably ralph looping a binary locally first is going to be best to get 100% recovery of types and function behaviors then letting a smarter model churn the final steps.
Fable is way too expensive for basic software stuff. Other models are more than good enough for that.
In general, if Fable isn't blocking you, there's a high chance a lower tier model would work fine.
The only problems Opus struggles with, Fable won't take on. I was porting some software from Win32 to linux. Opus was running in circles. Fable was going great until it saw some authentication code and bailed.
Most of the work I've done with pgrust hasn't had issues with Fable. The only time I've had issues is when building a fuzz tester to find bugs
When fable first released it was almost useless. Since then, it's improved a lot. It has been working on my binary ninja MCP server just fine. It flagged for cyber 1 time (no idea why), but it generally works fine.
I have noticed sometimes it likes to gaslight itself into thinking that everything its doing is allowed or allowable, I saw that it thought the game I was reverse engineering was running on a private server (it was not) so it assumed it had permission to do anything lol.
Nope. Always failed within 2-3 prompts. The most basic REST service you can imagine. Cookies are signed, that's crypto, banned. Completely useless model.
I have used it to write some scripts but it is incredibly expensive.
Nope.
And the lack of thought traces make it utterly useless.
> Claude Fable 5.1 follows explicit tool instructions reliably.
Moving stuff out the API into prompt engineering is obviously less reliable but necessary for progression to 'actual intelligence'. Will be interesting to see if it really is solid.
the counterbalance to the AI doomers has always been the fact that everyone has equal access to AI. i hate this new world where Anthropic believe they should be the ones to decide who gets access to super intelligence and who doesn't.
Thank the heavens for quota resets
I am not sure if Fable is worth it, at least with version 5 vs Opus 5. Opus beats Fable in quite a few benchmarks and at twice the cost I just haven't seen it provide noticeably better results compared to Opus. Has anyone noticed big differences? I did notice Opus maybe making more mistakes repeatedly but I don't have hard numbers on this. I hope Fable 5.1 brings noticeable improvements. I am giving it a go now on my 20x Max plan on a problem that Opus 5 has struggled for more than week now and has made very slow progress with regular regressions on the way.
My impression is that Opus 5 can be very impressive if you don't care about maintenance, novel-length comments, and really having any input in general. But otherwise it's borderline-to-totally unusable. It seems tailor-made to not have a human in the loop.
Opus 5 is better than Fable 5 except for creative programming work (like graphics). Fable 5 might be slightly better but the token cost isn't worth it.
How are you evaluating the models?
On the Fable 5.2 eval summary, Opus 5 only beats Fable on SWE-bench multilingual and multimodal.
I primarily use the models via interactive sessions enhanced with custom tools and skill. For that Opus 5's benchmark superiority has not materialized into greater productivity and frankly has been quite a let down.
The outputs are too often unreadable even after adding recommended prompts. There is an ongoing problem with the heron_brook system prompt affecting orchestration. [1]
I've used Opus 4.8 since the second week Opus 5 was released.
Over this time, Fable 5 has been reliably fantastic. Both in planning and direct execution on complex changes across code and infra.
I'm a bit surprised that there doesn't (seem) to be a section discussing ~performance across different modalities. This system card and blog post too-often default to an API-based use case when the gander primarily experience Anthropic's models via interactive sessions.
I understand waiting to comment until Opus 5.1 is available and handles these problems, though I am hopeful that Anthropic will confront the elephant in the room on Opus 5's failure to delivery great interactive sessions and the widespread negative feedback on the release.
It would show the org is paying attention, taking steps to balance model evals between interactive and API use. Also, some empathy for customers that wasted time trying to make opus 5 work for them.
[1] https://github.com/anthropics/claude-code/issues/80988
If you're doing something cutting edge like math or formally verifying algorithms, Opus 5 is a steaming pile of shit compared to Fable 5 and Sol 4.6, it makes countless stupid mistakes and is essentially incapable of completing the task without extreme hand-holding.
Reads like AI slop, surprised they can’t see it in their blog post. No human wants to read in such prose
While haiku is almost one year old. What a joke
do you think we'll go a full year without a new haiku lol
With compute crunches and everything I am not sure it makes sense for anthropic to commit to haiku as an endpoint and thus a product. There is no telling they aren't using a similarly sized model behind their existing opus/fable endpoints for various subagent / summary purposes of course.
They really should launch a new Haiku to compete with Luna imho. Luna is insanely good for the cost and it's my go-to for high volume batch tasks now.
So bullshit safeguards are still there.