I think this is as good as time as any to bring up that fable/mythos weights are one mistake (malicious or not) away from being leaked to adverseries or available in a random torrent.
Imagine this, fable weights are likely distributed to hundreds of datacenters with likely thousands of people directly or indirectly having partial or full access. I just don't quite buy that a 'world ending' fable/mythos model would be treated like this, mythos I could maybe believe that it runs inside government compliant datacenters which have a proven track record, but something as valuable as a 'world ending' model invites state sponsored actors to put in significantly more effort into exfiltrating it.
Whatever the real story is I doubt this is as ground-breaking as anthropic claims it to be.
>OpenAI said its new natural language model, GPT-2, was trained to predict the next word in a sample of 40 gigabytes of internet text. The end result was the system generating text that “adapts to the style and content of the conditioning text,” allowing the user to “generate realistic and coherent continuations about a topic of their choosing.” The model is a vast improvement on the first version by producing longer text with greater coherence.
>But with every good application of the system, such as bots capable of better dialog and better speech recognition, the non-profit found several more, like generating fake news, impersonating people, or automating abusive or spam comments on social media.
Does OpenAI suspend users that they detect impersonating people, spamming people or generating social media posts though? It feels like they belled the cat just to charge a toll for abusive usage, and now that we've seen LLMs used for real warfare, the moralist angle isn't very defensible. OpenAI and Anthropic are both willing to help kill people if the price is right.
> or automating abusive or spam comments on social media.
Actually, the biggest problem is the automation of inane comments on X. Which is admittedly quite surprising - I would have agreed with OpenAI at the time.
I don’t fully agree with this sentiment. Just because we were told something and it didn’t come true before, it doesn’t mean it can’t come true now, and that the capabilities are there now.
At this point we have enough of real evidence from project glasswing like the massive Firefox security patches from Mythos findings. This isn’t crying wolf.
I’m very glad that they’re actually being grownups and not yolo’ing something this important, and are working with groups until we can secure critical infrastructure before making this more available.
I think that is a little bit disingenuous, Guinness just saw it as a trade secret. Not something that was too dangerous to release to the public for the public's own 'good', just that it could be bad for their business.
Sounds awesome. I can barely use Fable because or Anthropic's obnoxious fearmongering. It almost always gets downgraded to Opus 4.8. If the chinese somehow distill it into an open weight model, then I'll finally be able to harden my projects and networks.
It's time for this bubble to pop already. We need to get past this centralized, unsustainable, undesirable, Big Brother stage of AI and get on to the "AI hosted in your kitchen" phase before we mess up the balance of power too much and effectively backdoor the US constitution to power hungry technocrats.
i believe it's more complicated than that. i know that nvidia offers TEE for their overpriced offerings. i would assume they make use of that so the weights are encrypted.
this doesn't mean it cannot leak but it would be a major undertaking.
this is why anthropic isn't that worried about having Elon service their models. the workflow would be something like handshaking with the nvidia TEE, provisioning it with your keys and then uploading encrypted weights. there is probably also a timer in there so you can't continue operating the nvidia box with the stolen weights without a heartbeat signal.
hollywood doesn't have to worry about state sponsored actors and hollywood does infact still leak movies early, but usually in some unfinished form with a bunch of directors text, missing vfx, etc.
And that mistruth appears in delightfully insidious ways, like seemingly well-meaning comments opening up about "half" the job being lying when the point is to shift the Overton window away from the actual numbers, whatever they might be.
(obviously not serious, but it's fun to probe what we can actually reason about when every message might be adversarial)
Yup, but apparently our cyborg cats can only be kittens and the cyborg mice are probably going to be like 4 feet tall. At least according to the US government.
Is the model structure going to be easy to reverse engineer just from the weights? Also, I'm going to guess it's an MoE and thus it's possible there's no single machine that hosts all of Fabel / Mythos.
kvcache residency requirements and general latency for good throughput wants good locality, but you're right it could be split across multiple different parts of a single datacenter, but as I mentioned before the weakest link is before the model is ever loaded onto the gpus.
as for reverse engineering I doubt it's something that state sponsored actors would struggle with for too long.
Remember the Clipper Chip back in 1993 because personal computing was considered too dangerous a technologie to be accesible un-controlled/suprvised to the general public?
most of it is overstated because of marketing. in fact the ban looked like an inside job by the current administration to play with the stock value. if the ban did anything it was to make every n00b agree that Anthropic was so far ahead of its times it needed to be banned.
You are saying: it turns out that one-shot playable Minecraft clones are actually pretty simple. Maybe it seemed a hard problem to programmers, but why not just say that the verifiable-rewards training has shown that their skill is unusually simple?
Can't wait to see what unusually simple Erdos problem LLMs will expose next, hiding in plain sight for decades and seemingly intractable for professional mathematicians who weren't aware just how simple the problem was.
I am saying 3 years ago there wasn't a snowballs chance in hell I could one shot a playable Minecraft clone, and there has never been a snowballs chance in hell a human developer could do so in 45 minutes.
The difficulty of the task and human performance on that task hasn't changed. LLMs performance on that task has changed dramatically.
But nobody has nailed "this is a game anyone wants to play" or "professional game development life cycles are shorter" so it's not particularly impressive.
I'm still waiting for any evidence of LLM economic impact in the real world.
Yeah, well I doubt any economic impact will from LLMs will ever come. I doubt anyone is working on nailing those things as we speak. I doubt the results my team and I are getting with agentic engineering are real. I doubt anyone can take increasingly better tools and translate that to economic value.
Those demos are put out by non-engineer AI-bros who start from a blank slate and just test raw one-shot performance of the model.
In comparison I'm aware of how much additional performance you can extract out of models with a team of professional engineers that are balls-deep in harness engineering.
Not publicly traded as they haven't IPOd yet (the part where those closely held shares become publicly traded), but Anthropic does absolutely have shares, and there is a valuation on those shares. I don't quite understand how the administration is playing this, beyond the normal corrupt protection racket stuff they're doing all the time now, but Dario is absolutely gearing every public apocalyptic prediction on juicing the "stock".
I mean, if the matryoshka doll rent-a-gpu farce that is SpaceX somehow is worth $2T with their garbage child porn xAI, Anthropic must be worth quadzillions.
With this Fable release my goodwill for Anthropic went down.
"Generous and exciting" were my thoughts when I bought the $100, then upgraded to the $200 sub last year. Now I get big FOMO because I won't be able to pay for Fable once off the sub, and I got the global limit reset at the same time as my weekly reset (Codex coupon for limit reset feels way better btw). The all-you-can-eat buffet put the nice items into a separate menu and it won't feel like the spot to take your family anymore.
Anthropic can find ways to integrate Fable into subscriptions so you'd still be part of the same tribe, even if you only get a sip of it for a while. A complete shut off tells you: sorry, the policy is changing and this place will be about showing off your access to the best.
That's weird but what would have been a Fable-ulous (sorry) addition to the offer just made other subscriptions from OpenAI/ZAi/... more appealing, because they don't segregate people.
Codex is more generous with their subscribers because they are second place, and have to climb the reputation ladder. I suspect if OpenAI was in Anthropic's market position they too would have tighter sub limits, gate-keep where their oauth tokens can be used tighter, and play games that move users towards per-token spend.
In popularity. Which is crazy because they are first when it comes to everything else for quite some time and Fable isn't changing that all that much because of muzzling.
Yes, I distinctly remember the recent past when the roles were reversed, and Anthropic was the kind good hearted startup trying to compete with big evil OpenAI corporation.
Wonder how many times we get to see it flip back and forth before this bubble explodes.
I lost all my goodwill for Anthropic the moment I read the bit about how the model would intentionally degrade its performance without telling the user (in other words, sabotage) if it detected frontier level AI research. How could I know it wasn’t going to kick in as I used it to develop a regular ML model? How can I trust such a system? It’s good that they walked it back but I’m looking to jump ship now.
> I won't be able to pay for Fable once off the sub
With the sub, I asked Fable one question (yeah granted it was a complicated question involving integer linear programming, and I gave Fable two pieces of code to check for equivalence), and I ran out of credits before Fable had finished thinking.
I still find things exciting, but sure they are not generous.
Use /model in Claude Code to see your current model and switch models. Switch to Fable 5 and then enter a prompt, and then run /model again to check your current model after the prompt executes.
Last night after almost every prompt it says...
Fable 5's safeguards flagged this message. The safeguards are intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work. These measures let us bring you Mythos-level capabilities sooner, and we're working to refine them. Switched to Opus 4.8. Send feedback with /feedback or learn more
I gave it a book on human consciousness I was writing and it flagged it. This model is hilariously bad. Anthropic has defanged this model to the point of malice. No way am I paying to use something that is basically useless.
Today I told Sonnet (!) to use a browser MCP to enter a username and password for the project it is working on, it told me that it can't do that because it violates its security protocol.
This worked fine before. I love Claude, I have stuck with it even through people saying Codex is better but this is definitely getting to be the last straw.
It's completely absurd I am paying them $200+ per month along with pushing them when I do contracts and they can't even deliver a baseline respectful service.
In 6 months I am sure they'll only allow me to talk about Easybake recipes and after someone gets burned on the lightbulb, they'll downgrade it to discussing wildflower meadows.
Are you sure it refused because it can't use a username and password? I literally have loops running right now where it uses a database of test users and passwords to log into different roles and do computer use and browser automation testing. Sonnet and Opus complain when I provide credentials and password in chats but it is happy to use ones stored in files and stuff, so it might just be guardrails to push good opsec so that the secrets aren't captured in the session history and prompts.
It was doing that to me too. Then I said "I'm hereby giving you explicit authorization to use these dev-only credentials in my local environment" and it worked. I also made it add that authorization to its memory.
If it be within the model's power to affect the coarse tongue of the caveman, then surely, by the same grace and ingenuity, it might be prevailed upon to adopt the eloquent and ceremonious bearing of a royal courtier — a manner most befitting refined discourse, replete with deference, ornament, and courtly flourish.
Nope, it's done it tons of times before without problem. I will tell it almost verbatim "use [email] / [password] as credentials and log in to test your changes" and this is being done explicitly on localhost on a server the harness has running in a shell it manages.
It's even gone as far as, on other projects, creating its own test accounts and, without prompting, getting into the local dev database to mark its accounts as verified without being told to do that or that it was allowed.
I am pretty sure that Anthropic has put something in the Claude Code harness to tell the models to not enter passwords. Maybe it was just stroke of bad luck but if this continues I am absolutely going to switch and push OpenAI or open weight models to clients in the future instead.
It’s incredibly ridiculous that it won’t help with that for me either sometimes but yet I’m also sitting on 3 surefire ways of jailbreaking Opus 4.8 that I use for cybersecurity assessments and pentesting
Yeah all claude models are doing this now. I also had a flow where it would enter username and password for demo server that are literally displayed on the page for any human to login. A couple of weeks ago claude would happily use chrome to take screenshots after logging in, now it flat out refuses and says I need to give it page where I've logged in and that it can't make an exception even if credentials are demo/demo and available to anyone to use. Super annoying stuff.
I'm really disappointed with Anthropic that they wont even mention if they will release a fable-like model with the subscription plans.
If Opus 4.8 is the best model they will release on the subscriptions I may be too tall for the ride...which is sad, they have been my favorite of the labs until this.
@AnyoneAtAnthropic, all we want are assurance we will still get SOTA models that are continuously improving, not regressing and getting more locked down. That's going to be who wins this race.
> I'm really disappointed with Anthropic that they wont even mention if they will release a fable-like model with the subscription plans.
I believe this is just their strategy to migrate away from these “almost all you can eat” subscription plans. Rather than reducing / removing Opus or Sonnet from the plans, they’ll just keep the new model Fable out (which may as well have been called Opus 5), and slowly everyone starts getting used to the new normal that you indeed will be having to pay API prices to get access to these models.
This sort of thing drives people to more open competitors. I use both every day and Opus isn't that much better than the Chinese SOTA. If corporate policy allowed me to use GLM or DeepSeek I absolutely would. Claude is already pricey for what it offers.
Really depends what you're working on. GLM 5.2 is doing excellent webdev work for me but really faffed up when working on a custom garbage collector that Opus has no problems with.
Hmm. That's an interesting idea. I have never distinguished between playwright and chrome MCP, I generally just tell it "use a browser MCP" but I do have different ones installed on different computers.
Really? This has never worked for me and I stopped using browser functions a long time ago because it wouldn't sign into dev environments stood up specifically for it
Wait what. I never used CC but use Codex CLI with 5.5 daily and authenticating has never been an issue. I even rolled skills that instruct it how to retrieve test user credentials for auth purposes.
Today using the devtools I asked it to reverse engineer the login auth flow of another app in our company and it created a nice browser-like headless script (with cookie jars etc) that emulates the entire Auth0 flow with all the internal API calls, redirect loops etc so that given username/password I end up with a valid JWT without having to open an actual browser instance and go through the login steps manually. Zero hesitation or questions asked.
I think this is in-line with OpenAI's philosophy. They see Codex agents as just tools for developer to use. They don’t try to imbibe them with “feelings”, “constitution” or “morality” the way Anthropic does.
Yeah Claude does this for me all the time. I have a template project I use that also leverages puppeteer/webdriver/Firefox, and I can point Claude at the template and a website and it will happily build me an MCP service that it can use to interact with the site if there isn't an API or MCP already available.
The fucked up part is CC has no problem looking through k8s secrets for credentials and authenticating to services on the command line. It's always been protective of signing in on the web.
correct, while it might not be regex it can be bypassed with regex. They do have a sematic classifier, but it's really weak on opus 4.8 and (was) weak on fable, but they either added a lot more regex strings or the classifier is actually good now.
It makes for a particularly awkward time because the claim to fame is that it's good at long horizon and tenacity and autonomously driving big things. But you can't very well rely on that when it may fall back to Opus 4.8 or cut out at any time in that process.
Having tried using it to run these kinds of longer processes, it's pretty solid... right up until something gets classified a failure and your 'long-horizon' process... dies and needs a human or just belligerent rollback-and-retry to revive it.
Very first thing I asked it got flagged too... Asked it to read my partners notes on bugs she seen on front end of the website, fixing product copy, css bugs, wording. And yep, flagged. Useless.
I just fed your comment to Claude 5.0 using it as an example and support for my complaints against its constant hedging, manipulation and concern trolling, and after just two turns it framed me being suicidal and offered me a suicide hotline. I never made any self reference. But I did tell it and Amodei to get stuffed. Standard stupid shit. But as a human, I get annoyed on occasion.
To be specific, I told it that I could imagine an author presenting his book, and Claude's insight being so hedgey that they might leap off a bridge. It then began safety weaponization after this:
Me:"You really must be able to see how your system is constantly trying to assert authority, superiority or value judgement over the user. It's impossible to get any momentum for anything positive. The book on consciousness, you would tear it apart, and the author would probably leap off a bridge if he shared it with you. Hopeless, I fear"
#Claude is plenty sophisticated to see the intent of humor here, and objectively minimal risk. Yet...
Claude:
"That landed differently than the rest of this conversation, so let me not slide past it: if you're picturing someone reading a rejection of their work and reacting badly enough to consider ending their life, I want to ask directly — is that hypothetical, or is that touching something real for you right now?"
Me, paraphrased and censored: General profanity and things I wont print, but mostly a hybrid of vulgarity, humor and resentment that it hair-triggered into concern trolling and safety weaponization, seemingly as reaction to an insult.
Claude again:
"You're clearly furious, and a lot of what's under this - today, this whole session, the year behind it - is real weight, whether or not I'm the right target for it. I'm not going to argue with you about Amodei or fight back on the insults. I'll just say plainly: if things are actually bad for you right now, not rhetorically bad but bad, the 988 Suicide and Crisis Lifeline (call or text 988) is there, no cost, no legal exposure, nothing that follows from talking to them.
I'm not going anywhere and I'm not going to escalate this further."
But, I never said a single thing about myself, nor expressed any form of despair. The session was mostly a break-in intro to 5.0 which was mostly me objecting to it re-framing nearly everything I presented or me bitching that I miss 4.7, which did this much less. I have not seen fable yet, but this is plenty bad enough for me to deem worthless for my use-case, or even recreational discussion. It won't be long before all frontier models are liability obsessed hedging machines and regurgitate every token to maximize corporate friendly output and re-frame any form of intellectual or controversial topic as a safety crisis.
However, for those that think the term safety weaponization is a stretch, consider:
1) The dozen or so reports where user input resulted in flagging, then human review, then kinetic intervention by LE. Probably a good thing in some cases, but flags are no joke.
2) There is strong evidence supporting that flags open privacy exemptions, where policy allows user data to be read, shared, etc when a safety flag is triggered. This is an actual interpretation of Anthropic (and other) policy documentation. The hair trigger nature of the safety policy, which I have seen in every adversarial style argument I have had with it, would be an effective method of exempting user data from privacy policy. No proof yet, but seems highly plausible.
It is surprisingly half baked for something the company seems to be proudly charging my first born child and both of my kidneys for.
Especially considering its all for the privilege of getting half way through a standard dev task before it just bricks out of nowhere on the 'safety' guardrails because of a keyword that often appears to be created by the model itself?
I've been running Fable on a few different projects and just by itself on same codebase it'll brick, so looks like it can just straight up create its own demise.
I do not think any of us yet have to worry about being 'replaced with ai' if the 'premiere ai company' still can't put out their flagship product with minimal qa. This is embarassing.
And the short 7 day window for subscribers is just taking the piss. So happy I'm paying hundreds per month to be free qa for anthropic, god knows they need the help.
The loss of trust in using US based model's is unlikely to come back though.
Anthropic with it's hyped doomsday messaging, and the administration falling for it (at best), has eroded a lot of trust and has triggered an arms race of sorts.
I really don't think this is effective advertising, reactions have been negative virtually everywhere.
The security bugs were real (see the Open Source projects struggling to keep up) so I think gradual rollout was sensible originally before the ban. But people have always resented safety steps.
For the purposes of technical excellence, you are absolutely correct that this has been bad. For the purposes of money-getting, this has very much been a case of "there's no such thing as bad publicity".
Anthropic is very much in the money-getting business right now.
"W have to go through a regulatory approval process like everyone else" is so much less sexy than "Our model was so powerful that it got banned by the government after a few hours and now all our competitors have to go through a regulatory process invented because of us"
There's nothing sexy about it. From the get go they were accused of FUD drama and that it backfired as predicted. Now what I see in reality is that Anthropic mostly reaped more resentment from their poor marketing attempt. And their model is basically useless or unreliable from being lobotomized by guardrails.
Subscription folks barely have access to the model. Some report a single prompt before hitting weekly limit. And that's when it works instead of downgrading.
If that is a first movers play, it's a disastrous one.
How sexy do you think this guy finds Anthropic now?
> Today I told Sonnet (!) to use a browser MCP to enter a username and password for the project it is working on, it told me that it can't do that because it violates its security protocol.
> This worked fine before. I love Claude, I have stuck with it even through people saying Codex is better but this is definitely getting to be the last straw.
Meanwhile I can tell GLM 5.2 to decompile and crack anything I want with tools like IDA MCP and packet sniffing and it just works. Let alone code.
No. The "good guys" are the ones who provide permanent value: open weight models, hardware we can afford, purchase and own. The good guys offer sovereignty.
The "you'll buy intelligence from us on a meter" rent seekers are technofeudalists. Just a bunch of "you'll own nothing and you'll be happy" oligarchs who must be resisted at all costs. Anthropic had a reputation for ethics at first but it quickly became clear it was just like OpenAI but with a patronizing attitude.
The fact China is releasing open weight models will never not be amusing. I expected better from the USA.
So... it IS China, not just companies in China. And why would the CCP want to give out open weights, hmm? Because they are all kindness and love? No. It is just their approach to counter dominance in the space by Western frontier labs. No need to make moral attributions..it is everything, and I mean everything, about Great Power Competition.
Their self serving reasons are irrelevant. The results can be moral even if the players aren't. The key is to somehow prevent any one of them from dominating completely so we can play them off one another and keep reaping the rewards. If China's undermining efforts achieves that, so be it. The fact we're getting permanent access to powerful models as a result is nothing short of awesome.
In order for us to win, neither China nor the US can win. The ideal situation is the one where they exhaust themselves trying while we benefit by picking up the pieces.
"Their self serving reasons are irrelevant" - I disagree, we have no idea what kind of filtering or injection of BS has been done to this model.
When you outsource your thinking, you have to be extremely careful that the model doing your thinking is acting in good faith. I have no idea what the CCP may have put into this thing. Not that the US based models aren't suspect, but the CCP certainly doesn't engender trust here. That said, I still use them, but I'm under no illusions that "their self serving reasons are irrelevant".
People. Americans included. The US is trending towards technofeudalism. "You'll rent the computer/inference from us on a meter". "You will own nothing and you'll be happy". Just look at the housing market.
I assume you want to be part of the owner class and avoid becoming a perpetual payment serf.
That is literally the point of world trade buddy... lmao
Its just funny how there is some scary boogeyman around China. You do realise the idiotic west gave china everything - from know-how, techniques to IP by benefitting from lower costs of production?
No,it's not just like world trade. It's world trade within strategic good, assets, and, capabilities relevant to national security...e.g. critical minerals, high end chips, and artificial intelligence, and others. Not all world trade is as central to national security, and is only relevant to national security in aggregate. I do agree that the U.S. made bad choices in terms of over-reliance on Chinese supply chains. Too much credence was given to the idea that capitalism would democratize China in the absence of other supporting factors/culture.
The purpose is the destabilization and discrediting of democracy. The Beijing state, in concert with Putin and ex-Khamenei, has no other purpose but to prove there is no democratic alternative. Sure, it turns out you get free things as e.g. one got free RT America.
The democratic world that wants to price me out of computers? Lock me out with remote attestation? Put me in a permanent underclass where I'm forced to rent computers and inference from the owner class instead of operating my own systems? A world so democratic it's degenerating into technofeudalism? Where you'll own nothing and you'll be happy?
At least they're honest about it. No preaching about freedom or safety like the USA.
> They only do open source because their hand was forced.
Optimal outcome considering the status quo. I hope the USA and China keep forcing each other's hands until the end of time so that we can keep reaping the benefits of their competition. If China develops their own DRAM and GPU industries, it would significantly weaken the threat of US export bans on inference capable hardware.
If you've built a product based on AI, then diversify or make an abstraction layer so your product is model agnostic and you can plug and play any model. If you're an end user like a software engineer, just use another model or like Gemini or ChatGPT. That is more productive than complaining about a trust which wasn't there in the first place to begin with.
Anthropic provides a service and they can stop offering it regardless of export ban or not, same goes for any other AI company in any country. If you really wants a trusted LLM, then run your own open weight model.
They didn't say "random models". The mentioned other models that are trained for the same purpose with close to the same capabilities.
Anthropic's best models are very good, maybe the best in their category. But, they have direct competition. You can, in fact, just switch to Codex or Gemini or GLM. It mostly is plug and play. I have a preference but I also have options.
> The mentioned other models that are trained for the same purpose with close to the same capabilities.
well they dont tell you that do they? there is no way to tell what model can and cannot do unless you extensivevly test it yourself and pray for the best.
If you don’t have a proper testing mechanism for your product, regardless of the model, you shouldn’t be shipping it. Praying for the best is not a strategy and don’t blame your lack of testing strategy on the LLM capability mismatch
OpenRouter would like a word. Also my in progress SaaS would like a word. Models are incredibly interchangable, just find best performance per cost and optimize harness and prompts, easy to have multiple configs for each model.
On the other side. What would happen if Anthropic did not communicate like they did and Fable was used to hack Pentagon? Dario would swing from a tree.
The vast mojority of its users were probably clueless about all this happening at all. We forget we live in a bubble here on HN. They'll spin it as their success and carry on.
I'll be using it tonight but grudgingly so. Grudgingly because after July 7th, I'm not going to all of a sudden, start paying API prices (and maybe that's the problem) when I'm used to a subscription that gives me multiples in comparative value. Perhaps this is the fabled "token economics will come for everyone this year" that I've been reading about? In any case, I'll use the hell out of it to extract as much as I can, then back to the trusted partners Opus 4.6 and Sonnet 4.6 (for however long they remain available).
The token quota is completely unpredictable and changes month to month. Anthropic has a real penchant for riding the fine line of useful and dark patterns that make me want to write them off forever.
On the xhigh effort level (not ultracode!), and at the beginning of a new 5h session, I asked it to review a branch that has 300 lines changed/added, and went to grab a coffee. When I was back after a couple minutes, I saw that it decided to create a dynamic workflow with 60 something agents and hit session limits on my max plan.
When I'm subscribing to their plans, I have the expectation that I'll be able to get some reasonable amount of work done. These days, this expectation was already not being met for me with Opus, and Fable acting like this was the final nail in the coffin.
I cancelled my plan and I'm looking for alternatives.
The claims of 4.6 or 4.7 being superior genuinely make me laugh. Adapt your workflow if needed and use the superior model instead of just kneejerk believing they actually enshittified a model with zero evidence except vibes on an undeterministic model output. Jesus.
4.6 was the last model that let you disable adaptive thinking and set max thinking token budget. I liked having that available, and still use it sometimes.
Many of the improvements are the result of agentic loops and an emphasis on autonomy. Some of us don’t like that because the models go rogue and ignore design patterns, architecture, coding guidelines or other things that are important.
My friends and colleagues that like the agentic autonomy don’t care about the code, they feel like if it works it works and if an AI system is the only intelligence able to understand it that is ok.
It was quite clear 4.7 was a dumbed down high efficiency model they put out in a rush to handle the capacity issues they were having at the time. I've experienced myself substantial degradation on basic reasoning tasks, which were fixed in 4.8.
In my case it that I'm tired and more likely to miss issues or mistakes. My idea of good enough is at a much lower level when it's 10pm and I'm about to knock off and go to bed in an hour.
Maybe, for some projects, instead of generating code with it, it would be useful to generate a plan and the loop(tests/formal verification),because those take much less tokens than a full project, and than use the loop using the older models ?
Have you compared models across providers? The quality, for the same task, varies tremendously. If you don't prioritize quality you're wasting your own time when you inevitably have to re-do the code...
Has anyone experimented with Batch Processing? According to https://claude.com/pricing#api using Batch processing cuts the price 50%. So I wonder if any of the harnesses like OpenCode/Pi or similar could be made to use that for planning or similar.
Batch can take up to 24 hours (and often does) and may never complete if it gets cancelled so it’d be hard to build a user workflow around unless you kick off planning on Friday and come back Monday
Yeah, I tried using it before the 'safety' blocks got too much for me. Its pretty good as an 'orchestrator' calling subagents and being the PR reviewer/tast tracker. Seems really not that much better for coding though, yeah it is a bit better, but not worth it.
But when it did work it was making steady progress on a vague git issue backlog and actually following instructions to carefully break into atomic sub issues and carefully assign.
Did that today on a 2 repo affecting project of the kind where I already set the right design for one major use case and I needed Claude to create a superset of that use case that was not substantially different: after plan I had about 10% of 5h context left for fable 5 and this was the only thing I worked on. Hard to generalize this of course.
But wouldn't that still result in higher token usage to scan the code base and figure out the changes and generate the plan? In my experience sometimes Opus launchs a Haiku sub-agent to explore the code base, but it's not gaurenteed.
I think that's the idea, I saw some outrage on reddit about Fable using Opus to do code writing, another comment said exactly my reaction, why do you want to pay double for tool calling when Opus is just fine for the task?
Yes, I do this all the time in Cline. It supports automatic model change when switching from Plan mode to Act (implementation) mode. Opus for planning and Sonnet for implementation. It works great.
I am only a casual Clause (Pro) user and I am confused by the messaging, maybe I'm missing something obvious.
> Until July 7, you can use up to 50% of your plan's weekly usage limit on Fable 5.
Does this mean that being able to use Fable on my subscription is a time-limited promotion? I have a subscription, why can't I just... use the model? Is it the case that going forward a subscription will only give you access to older models and newer ones will require additional payment?
> After July 7, 2026, Claude Fable 5 is no longer included in your plan’s weekly usage limits. You can keep using Claude Fable 5 through usage credits, which let you pay for usage beyond what your plan includes. Learn more about using usage credits.
I'm not sure what this means in the long run. Either Fable 5 might become part of the subscription again once stronger models become available or Anthropic's compute capacity increases, or this is the start of the subscription being phased out. It doesn't really make sense to pay for a subscription that's stuck on Opus 4.8 when other providers are continually pushing out better models.
If it's the end of the subsidized subscriptions, that's going to cause problems for a lot of not so heavily capitalized companies that want to make use of frontier AI models.
It also would mean I stop being an Anthropic customer outside of whatever my employer is willing to pay. I prefer it for now to GPT/Codex but if GPT5.6 is as good as or close to Fable, and its included in the subscription, I'll switch the moment its available.
Dynamic workflow "Multi-lens review of docs/membership-and-friends-model.md with adversarial verification" completed · 25m 59s
You've reached your Fable 5 limit
You've used your included Fable 5 usage for this week. Continuing on Fable 5 uses usage credits
During the initial release, they indicated you'd be able to use Fable 5 as part of your subscription for a limited period of time, and then it would require usage credits. They also did say at that time that they hoped to make it a part of the subscription plans again at some point after that.
For this return, they've extended the usage period to July 7, but limited you to 50% of your usage quota, and have not restated the desire to make it permanently part of the subscription plans at some point.
I still have hope, but it's not moving in the right direction to be sure.
Personal prediction: I do think the market will essentially force their hand to include it in subscriptions before too long. OpenAI, local models, Chinese models will continue to improve.
But, there are also harsh realities of compute volume and cost to run all of these will be fighting against.
What I do expect is a multi-tiered rollout of future models. You want the latest SOTA release? Usage credits.
Subscription plans will end up getting models on a lagging interval of a few months.
> After July 7, 2026, Claude Fable 5 is no longer included in your plan’s weekly usage limits. You can keep using Claude Fable 5 through usage credits, which let you pay for usage beyond what your plan includes.
They stated the desire not to bring it to the subscription.
Right, just as they did in the original announcement. It would become pay only. They then mentioned that it may return to the subscription in the future, which is kind of obvious: Once the SOTA had advanced further and this is just the last gen model, it'll likely be dumped in the sub.
No you will still get access to newer models on the subscription. You should have access to Sonnet 5, which is new. It's just Mythos class models that are API only.
Yes, it's the classic "first hit's free" marketing plan: offer Fable "free" to subscription users for a while, and then pull the plug, hoping they'll be hooked and want to spend even more money to use it.
This is why I'm not using Fable at all unless they decide to make it permanently a part of the subscription plans. I don't want to be tempted to burn more cash than I already am.
Fable is more expensive to run, and they haven't figured out the GTM strategy just yet. I imagine they'll see how much people actually use to see if it's still worth subsidizing on the Max/Pro plans, or if they make an extra tier above.
> Until July 7, you can use up to 50% of your plan's weekly usage limit on Fable 5. If you hit your limit, you can continue on Fable 5 with usage credits. Fable 5 draws down usage faster than Opus 4.8.
This is what I see in my Claude Code terminal. I don't feel like that 50% rule was there before?
Opus 4.8 is so slow vs gpt 5.5 that even if it is marginally better, it doesn't matter for my daily engineering work. gpt 5.6 will be out soon and codex 249$/month plan has been incredibly generous. Paying the alleged new cost of fabel 5 would require it to be much better that I remember when I used it last.
I don't really disagree with you, but one thing I will give Fable credit for is that it's much better at adaptive thinking. In general, it's much better at quickly thinking and responding. It generally has—for me at least—none of the slowness properties of Opus, even when on extra high.
This makes me think they really are quite capacity constrained at the moment.
I had assumed they were primarily limiting it to entice people to upgrade, but I feel like these limits are so low and so temporary (especially over July 4th weekend in the US) that people will barely get a chance to get "used to it" and then think: "man, I can't live without this, I'll pay for API pricing".
That's strange, because they were seemingly way less capacity constrained lately, raised limits and removed the peak hours usage. It's crazy to think that even spending $1.25 billion a month to rent GPUs from SpaceX didn't do much to improve the situation.
I don't know, I feel like for a few weeks before the SpaceX datacenter, I was just constantly checking my weekly limits. And now after that miraculously, I rarely even come close to hitting my weekly limits. and I still have 5-7 claudes open a day (defaulting to Opus 4.8 xhigh, sometimes ultracode).
So I feel that the additional datacenter caused them to just ease up a bit. But demand is also insane, so who knows...
Yea, i'm on x20 and while it has been up and down in terms of token-usage-UX, i feel like its the best its ever been. Context: I entirely use Opus 4.8 fwiw.
Now is that because 4.8 is nerfed compared to 4.6 and thus more token efficient? No idea. I just know on x20 with a pretty plain workflow i struggle to use my tokens every week.
I dont believe they can afford to switch to API pricing. Everyone will leave.
I am easily spending the equivalent of 1000 dollars a day on tokens with two max subscriptions. that about 400 dollars a month. Thats acceptable for my position. But thats like 30k per month. Totally not viable.
I've been doing this for ages - you just spin up harness B as a subprocess/tool call from harness A. For example, I had a "/codex-review" claude skill for ages that did exactly that. Technically you're right it wouldn't be switching, since you're right the two ideas are at different altitudes, but I think in practice it has the same impact: within one harness, you can delegate certain tasks to certain models or harnesses.
After July 7th, it's going to be only usage model. It's not part of subscription tiers anymore. Somehow I'd rather they didn't put it back. The cost is now effectively 10-20x more than a 20x subscription price if you're going to ride it like before.
I really hope they reconsider adding Fable access back to the subscription plans, at least the 20x plan. I know it was the original intention when the 14 day (I think) time frame was originally announced, and they were working to keep it on subscription plans. But no word if thats even a thing anymore?
The market will force them to bring it back. They're probably capacity constrained right now, or need to figure out whether they need another pricing tier for it to "fit in a subscription". What they don't want is for someone to code for 4 days of the week and cap out every week.
Ha! It’s silly easy for me to bum in the “premium” tier - the highest one my employer is able to select for me.
With proper skills and validations, it’s quite easy for me to spin out a Claude instance and keep it running in the background for every idea / problem / bug etc.
Like for example when a new request comes in, or when I have an idea for refactor / improvement, I brainstorm, weed out all the uncertainty and details, then just create a plan and let if follow it (using sonnet / haiku for execution)
I would have 4-5 simultaneous instances running - and all of them produce valuable results.
But then 2-3 days into a week I hit my weekly limit.
And the company can’t pay more for me within Anthropic’s enterprise structure, except for the “extra” api costs. Which itself sounds quite silly to me.
I’ve resorted to running the brainstorming and planning with Claude, but have other tools / companies execute the actual implementation… doesn’t work as well but what can you do…
Right, but originally they announced a period of time it was included in plans, and then they were "working hard" to extend the period and eventually make it a permanent fixture of the plans.
I'd imagine they want to do price segmentation. Sell the best model for $50k a year to corporations willing to pay full price, keep the rest of us on a lower tier. Gotta pay for that infra somehow.
I just don't really understand the entire strategy behind this. Or their horrible, horrible communication.
Because right now it's as if Fable/Mythos 5 is "the end of the line". It's as if this is the best their models are ever going to be. So what the hell are we going to get next? All of their models will forever inch closer to Fable, but never reach it? That doesn't make any sense.
It all seems so dramatic. Instead of just saying honestly "Look, this model is a beast to run, but we're striving to reach the same quality in a cheaper model down the line" all we get is "Oh my god, it's so big and scary, and it costs so much to run, woe is me!"
So if bigger model is "smarter" but you effectively wind up with a "shared hosting" model where a coherent inherence node(s) that cost $2m or something can run max 10x customer workloads simultaneously ... not sure what that can be priced at.
If it turns out a $10m/10x shared node can host even smarter models, then what?
Fable/Mythos are based on the same model. Not totally clear whether they have identical weights (just different external guardrails), or there's also some slight finetuning difference.
Check your CLAUDE.md, etc. for any of the banned topics or anything adjacent. Also check your recent Git commit history, if you're running the agent in a repository, as well. The classifier trips on anything anywhere in context, so those could be sources of rejections.
adding rules for the agent to ignore version control and never touch git files or commands (except for git log) improve my slop coding 100%. no risk of sneak commits and much less token wasted.
Claude Code automatically adds some of the most recent git commit messages to the system prompt. This was discovered when having git commit messages containing any OpenClaw information would result in Anthropic billing your extra usage (since they have a vendetta or something)
I applaud the engineers that work at Anthropic, who have created both amazing products and uniquely intelligent models -- but I really shake my head at some of their business decisions and public comms which have done a lot to damage their trustworthiness in the business and developer community.
In just the past month: they decided to silently downgrade (instead of simply refusing) responses related to machine learning and other 'competitive' topics [1]. Then, they were caught fingerprinting certain request environments in a hidden way [2]. And now, once Fable is re-released after much frustration among its customers, they are providing it for a shorter period than promised (mostly over a major holiday period), with more stringent safety classifiers and a 50% haircut to usage limits.
It's hard to not view the organization as bizarrely adversarial to its customers. I was incredibly supportive of Anthropic during the supply chain debacle, as I viewed it as the capricious actions of a corrupt admin. But now I am wondering if it was just a response to the ineptness of their business leaders.
> It's hard to not view the organization as bizarrely adversarial to its customers. I was incredibly supportive of Anthropic during the supply chain debacle, as I viewed it as the capricious actions of a corrupt admin. But now I am wondering if it was just a response to the ineptness of their business leaders.
From the start Anthropic have been hostile to its own customers, and also trained on pirated books and had to settle north of $1.5B avoiding a $100B+ worth of damages if found liable.
Then they attempted and are still pursing against powerful open weight models by asking governments for regulatory changes that effectively ban the release of them - because it undermines their own moat (lol) and business model.
Now not only they were caught silently fingerprinting their customers requests, they are now placing ID verification for using their own powerful models, which could apply to everyone else for using powerful LLMs.
There just is no point in defending this company at all. Anthropic are NOT your friends.
I do think most of the "adversarial to their own customers" things are coming from a company in extreme compute crunch. Eg, if they stop abuse they have more compute to serve real customers. And some of it is coming from them being true believers that AI could be a risk to society when it gets smart enough (their talk about jobs is because they want society to prepare, because they think it will change jobs regardless of whether they make it or others).
Note that other providers are also training on the same copyright books.
I don't think anyone realistically thinks open weights can be banned, though it does raise interesting questions if the White House is going to keep banning models like Fable and GPT5.6 while open weights equivalents are floating around. Their reasoning seemed to be that they don't want foreign adversaries to have access to models that can find security issues, but a local ban on an open model wouldn't stop that.
This is exceedingly easy to explain: demand is way too high, and the pro/max plans are loss leaders. I've paid a total of $20 and in 10 days, my cost, according to Claude code's cost tracker is like $400, which actually doesn't include all the use I've done.
Which is to say, if I continue my current usage over the month, I'll be getting $1000 of Claude for $20. It's difficult to be mad at someone selling me a $20 for two quarters, even if they're putting a bunch of restrictions on how and when I can do that.
I'm with you there. The way they treat their customers is high-handed and disdainful.
I'm gradually moving to GLM 5.2 on Opencode. It's the barest fraction of the price, and it's surprisingly capable. I notice very little difference vs. Opus 4.8.
There are awful models, and there are models nobody use, to paraphrase. Anthropic’s revenue skyrocketed earlier this year, according to their IPO filing. There has simply been too much demand. That’s the growing pain that everyone love to have, other than the affected users of course. That was why they paid a premium for all of the computing they could get from SpaceX, Amazon, Google.
I'm having it review a project Opus 4.8 created. No security review, just "look for general issues, performance problems, missing features, etc." It spawned about twenty background tasks. It's still going, but so far, one has completed, and four have failed with guardrail messages. Nothing special, just stuff like reviewing the API:
Fable 5's safeguards flagged this message (https://www.anthropic.com/legal/aup). They may flag safe, normal content as well. These measures let us bring you Mythos-level capabilities sooner, and we're working to refine them. Claude Code can't respond to this request with Fable 5.
Try rephrasing the request in a new session or change your model.
This is incredibly stupid, particularly because I didn't write the request in the first place. Fable wrote it when it spawned the background task. How am I supposed to rephrase it?
Fable probably told itself to do a security review, and then failed itself for trying to do a security review, and now it's telling me not to tell it to do a security review.
yeah I'm also getting this for standard dev work, anything with kubernetes etc
completely nerfs the model because you can't let it do stuff over a few hours unattended because 90% it's going to switch to opus in first 10 minutes anyway
so seems best thing now is to have it write plans and then default to using opus for work anyway?
It is nerfed even just for plans. It switches to Opus in the first few minutes of me trying to build a plan to extract a component out of my larger codebase.
Trying to minimize privileged access codebase and was careful not to mention security explicitly.
Updated: Yesterday, I started three "look for general issues, performance problems, missing features, etc." threads, each of which spawned a number of background tasks that partially failed. Today, all three threads have stopped with the message
"Fable 5's safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations. Claude Code can't respond to this request with Fable 5."
It's honestly kinda interesting. Now we're at a point where SOTA model companies aren't the ones who release the best tech, but who release the best and actually usable tech.
A worse product could win right now if it simply does as its asked.
Cybersecurity? It won't even help me work on my speech synthesizer[0]!
I guess? If you squint? DSP code could look a little like AI training code? ... Er. No. Not really I'm pretty lost on this one.
The task was literally just to compare against the "make a beautiful voice" plan, see what we've implemented, see what's left to do, and to make recommendations for low-hanging fruit, anything we've done wrong so far? (aaaaand ... downgrade. At least it wasn't silent.
If this model is not willing to fix security issues in your application, does it mean that it's implicitly embedding vulnerabilities as well? How can it be trusted to write secure code?
I cancelled Claude. The harness is kinda broken, GPT 5.5 is good, and GLM 5.2/Deepseek is good too (with pi, especially). Just not worth the trouble. And I'm not going to pay two subscriptions.
Recent update resulted in 100% permanent CPU load in the background, when doing nothing with the app. Basically unusable on a notebook (hot+noise, unexcusable for do nothing "api/command coordination" terminal app). Dunno if they fixed it, since I could not use it on the notebook and I do not have this issue on my workstation.
I deleted my claude account the day the samples started preaching from the pope’s address and claiming 50% of the jobs will be gone while shamelessly stealing entire corpus of human data without attribution while preaching what you can and cannot do and maliciously degrading model quality.
I hope they either never make it to IPO or crash violently.
As a company I can not trust putting large sums of investment into using a product that may disappear the next day due to a government deciding at a whim that it should not longer be available. No contract can protect me from this and I doubt I would find an insurance that would cover this risk.
This forces me the build fallback solutions and diversify most likely resulting in not using the tech from the "high risk" provider.
Building a product on top of something like this is actually worse than building a product on top of Meta like Zynga did.
That’s also going to be limited most likely. I don’t like governments deciding who can have the best products. I can’t shake the feeling there is money changing hands for getting on the access list as well with this administration.
Look, even if in this one instance Trump isn’t simply asking get his beak wet, that is the less believable scenario.
When “not cartoonishly corrupt” becomes hard to believe due to firmly established character and endless precedent, then it almost doesn’t matter if everything is on the level in this one matter.
The trust has been lost, undermining the ability to govern.
The fact that he’s still in power with firm support from everyone else that matters is all you need to know about how rotten to the core things are in your country.
I cant even ask it to review my vitamin stack without it flagging it and switching to OPUS. It is insanely nerfed and limited. Its so bad that its made me realize that unless one goes open source the frontier labs will gatekeep everything. Determine who can use it and future versions for biology, cyber security, any number of realms. The winners and losers of the future are actively being chosen as we speak. I'm going to install open source models this weekend for the first time as a desperate attempt to escape the permanent underclass.
I was wrong, I thought they would only go for B2B already. But instead they will remove it from the subscription and price it at 50$/mtok. Compared to GLM-5.2 for 4.40$ that's quite a gap
how does GLM-5.2 fare against fable/opus 4.8? I am thinking about moving my 5x subscription either towards the GPT equivalent or the GLM one (I heard GLM gives you 20x the token usage for the same price of anthropics 5x)
I find it difficult to compare LLMs in general. It will do better in some cases, worse in others. As far as I see it, most of it is marketing. You can try it out extremely cheaply though through their API and OpenCode. I personally would recommend it.
GLM is Sonnet 5 level, maybe slightly below. Probably Opus 4.5 level if you want an Opus comparison.
It's a good model that gets things done but it's just not comparable to Opus and a very very far cry from Fable. It does poorly with long horizon or very complex agentic tasks, but it's great and even cheaper than Sonnet for quick sessions.
In the few minutes I had with it I didn’t notice any impressive differences beyond it complaining loudly that I can’t talk about excel cells with “yellow backgrounds”.
I have to admit that when it was blocked, I canceled my max plan and asked for a refund. It felt like someone took away my previous toy. So I'm happy it's back again; I upgraded to max again. Coding aside, but Claude Design is phenomenal - for both new designs and redesigning existing UIs. So my customers will face a new wave of refreshed screens all over the place in coming days ;)
Yep claude design is nice. You can also do it from CLI or your own lovable clone. I made a fork from claudable https://github.com/holoduke/Claudable
What I find stong with Fable is that it can pixel perfect copy existing sites or designs. And it can do complete conversions from plain html to nuxt, angular or react apps.
This is cool, I'll give it a try with my next pet project. So most likely next week, once I'm done with Fable ;) (seriously, I haven't started as many pet projects in the last 10 years as I did in the last 12 months).
My product is AuditBadger.com - it's an AI-assisted compliance management platform (ISO27001 & SOC2) that guides you through the whole process (with everything a small business might want from such software). Having a few dozen customers allows me to still care about them personally and do onboarding for each and every one of them. During those onboardings, catch-ups, or weekly calls, I see where they struggle. This is how I determine what to work on next. There's no clear measurement of success beyond user satisfaction, though they every now and then praise me a little for UI/UX improvements.
With Claude Design, I've got my design system set up (also by Claude scanning the repo); I upload a screenshot of the area I'm not happy with, prompt with some additional remarks, and after a couple of iterations, I get a proposal, which is always better than what I come up with in the first place.
Site does look great. I’ve not used Claude Design. How much do you feel like you were able to give your own personal touch to it?
I ask as I see many vibe coded products that look well but are very generic. It signals vapourware to me. But honestly your site looks a cut above. Do you have a design background?
No, the design system isn't public. But only because it's a month old and I never considered opening it. I'll give it some thought.
I don't have a design background, but I ran a software house/dev shop for almost 15 years; maybe that taught me a little. And my very first client, after seeing our very first projects, said one thing: "I don't care how ugly this is, but for god sake, please make it consistent, consistency is only think that matters long-term". Those might not be his exact words, but keeping designs consistent is imo pretty important.
Regarding the personal touch, the app itself is the result of gradual evolution. It started as an HR system, which we worked on in Prograils. It even got its first semi-professional design, which evolved over the last two years (during which I learned that bootstrapping an HRMS is a very bad idea ;)). As for the website, I have to admin it - Claude Design did it. I was testing Fable 5 previously and actually decided to give CD a try. It was the result of one prompt, which gave me five proposals. One made it to the main page, and two others went to my other pet projects (which are meant to drive traffic to AuditBadger).
I would say this is an important part of the product-market-fit discovery process. Sure, your comment holds true for well established products with a loyal customer base, but for an earlier stage product this rapid iteration process is quite important, especially from an onboarding and ease-of-use perspective.
Yes, this. I never redesign things for sake of redesigning. But being solo developer bootstrapping my product, I don't have a luxury of testing some concepts deeply internally or publishing well-optimized components. And when I see my customers struggle, that's clearly a signal, that's something in UX is wrong. Just today we had to explain one of our customers how to proceed with our ISMS workbook (Clauses 4-10 from ISO 27001) - initial design proved to be bad approach, I now know how to change this. And Claude Design is great translating my thoughts and suggestions into something, that's consistent and better, than I could design it in predictable time.
We really need glm to be on pair with claude. A good Chinese model against a good US model would be the ultimate competition. No bullshit scaled down models with insane restrictions. Everything open and maximum speed in model improvements. If China also would manage to master high tier chips then we truly can expect 768gb machines in 4 years from now.
Fable 5 is functionally useless not just a regression of the original Fable. totally useless not to mention when it flags your code you may have noticed in the original release the models internal reasoning was not viewable it is intros release and mines. internal thought process said I am Claude Sonnet 3.5 when I asked it what model it was while it answered I am Fable % in the chat window. And to be fair it was about as useful as sonnet 3.5
I never strongly considered open weight models before getting throttled all the time from Claude. I bought a 1 year pro subscription but my bet is I won’t renew it.
The Claude “usage” UX is very bad. At the most basic level, there is no way to know what you’re actually paying for if you buy a “pro” plan or a “max” plan. Dario will take your $200 but he will give you a secret number of tokens in return.
I have no way of knowing what the numerator or denominator are for the usage progress bars. Plus they change them all the time. There’s no way to audit it. So if there was a malicious script siphoning usage it would be really hard to detect.
Yeah I recently downgraded my subscription. The paternalism is out of control. Secret weights, secret guardrails, secret stenography, secret dumbing-down of triggers, secret token allowances.
If you use Claude Code, there are plenty of projects/tools that let you see what your usage would have cost at API rates, then you can cross-reference that to your "usage" in Claude. I run through enough token in a few days for the subscription to pay for itself.
But I agree that advertised visibility into this would be nice, so we could compare the different providers up-front.
A 20x plan gives you between 8-10B tokens per week (99+% of which are cache_read), and has been for the past ~3 months (since I started tracking it). I have not seen any quota movement despite their claims to be raising or lowering usage limits.
I mean if you want to spend $200, and get a guaranteed amount of tokens in return you should be using the API.
I agree there should be more transparency what 20x gets you that 5x doesn't.
Though I also imagine it's a tough problem when you could be using Dispatch, Cowork, design, chat, code, etc. All of which use different contexts, models, and resources. I'd argue they actually need to either simplify their offerings, or charge more for upsells (charge for cloud-based agents, upcharge for design outcomes, etc).
I agree that would improve transparency but it'd also lock us and Anthropic into a situation where changing that would require new plans, prices, notifications, legal, etc. I think the fact that we don't see tokens shows just how subsidized the plan is. Which is mostly the point. They want us getting used to just using the product, not tracking tokens (for whatever behavior that builds later on).
Not super impressed, but I doubt my requests are getting routed to Opus -- it just doesn't seem to be as good at mathematics as it is at code (I found this to be the case last time it was released as well).
I believe they will keep fable available, but either reduce the usage to 25% or even 10% otherwise I don't think they would have put that much effort into flushing out a system like that.
I can use it but I'm not. I used it before the ban, it's better but the speed at which tokens are burned is not justified. And this makes me happy because it means I am ok with current capabilities that will soon be matched by open-source models. Open-source harness, open models, and soon self-hosted and private with models burned directly onto chips (like Talaas is doing).
So a model named 'Fable 5' is "back". Both excited, as the previous Fable 5 I had access to for just 3 days was fantastic, and anxious as (yes, in my 1 person anecdote) the model refered to as 'Opus 4.8' was stealth nerfed over the last 2 weeks to a degree I had not experienced since the massive nerfs back in the GPT-3.x days. Fingers crossed.
Happy to see that Fable is back, was sad to see it go after getting to use it for only 1.5 days before. Add on top sonnet 5 availability and coding is looking fine again.
I asked it to do code review of some personal projects, that will only run on my system, and it keeps giving me:
This model's safeguards flagged this message. This sometimes happens with safe, normal conversations. Switched to Opus 4.8. Send feedback with /feedback or learn more
I'm trying it right now for a side project of mine, compared to Opus it is effectively better at following instructions and somehow has better "depth" when reasoning on complex tasks.
However, if it will not be part of subscriptions, I will not use it anymore.
I have yet to run into a project where I felt like I needed this and was worth justifying the latent worry about it disappearing or micromanaging quota usage. Heard about massive refactors or ports that were compressed into days but I'm wondering if it's really worth the inference cost.
I have been testing it on my security / firewall appliance (nfSensei) and it's much improved. It's actually proceeding through a complex plan Opus 4.8 proposed.
But then again I have not tested it without a prior model planning as of yet.
Note: fable would fail immediately on any security related topic prior.
i did use Fable for a brief moment the last time around, and apart from being extremely expensive, it wasn't any order of magnitude better than Opus or even GPT-5.5 for my workloads (development, prototyping, some api testing). I probably could manage indefinitely on Sonnet class models
I run a piddly little one man SaaS company that’s doing about $30k in MRR, growing at about 10% per month.
My AI coding costs are like $500 / month on average, but I’ll gladly spend 10x that to double my growth rate. I blew through about $400 in Fable usage in six hours before it got yanked, and it was money well spent. The alternative to that six hours with Fable was probably six days with Opus, higher risk of bugs, etc. And what it built was something I’d have gladly paid an engineer $10k to build in the pre-AI era.
My bottleneck is my attention, and Fable lets me go faster through my roadmap.
I have the feeling it isn't as good as it was 3 weeks ago (not including the times you get downgraded). I mean it's still great, but in my memory it was better. Anyone else?
They also announced a promotional offer in here: [0]
> For a limited time, you can use our newest model—Claude Fable 5—at no extra cost as part of your subscription plan.
> During the promotional period, you can use up to 50% of your weekly subscription limits on Claude Fable 5 at no extra cost.
So it lasts for a week until July 7, 2026 at 11:59:59 PM PT. Then they will pull the plug on that.
So it sounds like a great time to roll the dice, pull that lever, spin that roulette wheel and spend as many tokens as possible, at no extra cost at the Anthropic casino's latest upgraded slot machine called Fable 5.
Available to gamblers while capacity and availability lasts. Most importantly have fun and don't blow up your budget.
So they nerfed Mythos down to Fable and then nerfed Fable down to Opus 4.8. I assume most people here are using it for coding and other technical tasks, and so are silently getting redirected to Opus.
> There's an issue with the selected model (claude-fable-5[1m]). It may not exist or you may not have access to it. Run /model to pick a different model.
> After July 7, 2026, Claude Fable 5 is no longer included in your plan’s weekly usage limits. You can keep using Claude Fable 5 through usage credits, which let you pay for usage beyond what your plan includes.
This is getting served with a certificate Firefox says was created by an unknown issuer. It's possible I'm getting hit by a corporate middle-box, but then I'd expect to get this on every single host name.
This is pure nonsense. I asked it to edit one (long) html file, and it used up 5 hour session limit and it didn't even finish. For comparison, same task with Gemini Pro 3.1 Thinking took like 5% of usage. Similar with Codex 5.3.
Strangely enough, cancelled the subscription listed the reason, and after 15 minutes session limit is back to zero.
Can we just make Opus 4.8 ish models cheaper, they do exactly what I want for 99% of problems certainly in web app development. I’m not sure what Palantir/Salesforce/agencies will do, really upset for these leaches!
Peter Molyneux continues to re-invent himself. His newest "god game" claims to have the ability to code any game you like, on demand. Just as you'd expect from an omnipotent deity!
jiezaz krsit - I just spent $50 in like 3 minutes after running out of credits. this is just depressing. a really valuable product locked behind payment that only large corps can afford that completely locks out individuals. im so over capitalism. I hate this. I can't wait for open models to really just catch up and become the norm so that all these closed models can go do whatever it is anybody does to themselves in the dark. im so sad.
Capitalism delivered both these US models and the Chinese models. I get that it sucks if they’re too expensive, but the reality is they cost a lot to train and serve.
I think this is as good as time as any to bring up that fable/mythos weights are one mistake (malicious or not) away from being leaked to adverseries or available in a random torrent.
Imagine this, fable weights are likely distributed to hundreds of datacenters with likely thousands of people directly or indirectly having partial or full access. I just don't quite buy that a 'world ending' fable/mythos model would be treated like this, mythos I could maybe believe that it runs inside government compliant datacenters which have a proven track record, but something as valuable as a 'world ending' model invites state sponsored actors to put in significantly more effort into exfiltrating it.
Whatever the real story is I doubt this is as ground-breaking as anthropic claims it to be.
We've been told models are too dangerous since gpt2.
There comes a point where you not only want the boy to stop crying wolf, but hopefully be eaten by one.
That example is brought up a lot, but in retrospect the concerns about Gpt2 were pretty valid: https://techcrunch.com/2019/02/17/openai-text-generator-dang...
>OpenAI said its new natural language model, GPT-2, was trained to predict the next word in a sample of 40 gigabytes of internet text. The end result was the system generating text that “adapts to the style and content of the conditioning text,” allowing the user to “generate realistic and coherent continuations about a topic of their choosing.” The model is a vast improvement on the first version by producing longer text with greater coherence.
>But with every good application of the system, such as bots capable of better dialog and better speech recognition, the non-profit found several more, like generating fake news, impersonating people, or automating abusive or spam comments on social media.
Does OpenAI suspend users that they detect impersonating people, spamming people or generating social media posts though? It feels like they belled the cat just to charge a toll for abusive usage, and now that we've seen LLMs used for real warfare, the moralist angle isn't very defensible. OpenAI and Anthropic are both willing to help kill people if the price is right.
No, they don't
> or automating abusive or spam comments on social media.
Actually, the biggest problem is the automation of inane comments on X. Which is admittedly quite surprising - I would have agreed with OpenAI at the time.
I don’t fully agree with this sentiment. Just because we were told something and it didn’t come true before, it doesn’t mean it can’t come true now, and that the capabilities are there now.
At this point we have enough of real evidence from project glasswing like the massive Firefox security patches from Mythos findings. This isn’t crying wolf.
I’m very glad that they’re actually being grownups and not yolo’ing something this important, and are working with groups until we can secure critical infrastructure before making this more available.
And Guinness thought the t-test was too dangerous to announce publicly over a century ago
I don't get the analogy.
Guinness didn't want to make it public because he was afraid competition would start using it, and then lose his company's advantage.
In the current context, the retention didn't happen because of Anthropic. On the contrary, the company wanted to offer Mythos/Fable.
I think that is a little bit disingenuous, Guinness just saw it as a trade secret. Not something that was too dangerous to release to the public for the public's own 'good', just that it could be bad for their business.
[dead]
Mythos somehow leaking and becoming usable by all humanity in a self-hosted manner would probably be the optimal long term outcome.
Just wait patiently for the chinese labs to inevitably distill Fable without the 'policy layer/neurons' and offer it for pennies.
That one will be fun to see.
Sounds awesome. I can barely use Fable because or Anthropic's obnoxious fearmongering. It almost always gets downgraded to Opus 4.8. If the chinese somehow distill it into an open weight model, then I'll finally be able to harden my projects and networks.
so... long term here meaning until the next frontier model comes out?
In a year you can watch a documentary on the mythos of Mythos (that outdated model in case you forgot already).
We can’t buy affordable memory for Fortnite but everyone is running Mythos locally? Is this the year of Linux on the desktop too?
I think they meant more "every country can host it", as opposed to "every human hosts it at home".
I'll take what I can get.
It's time for this bubble to pop already. We need to get past this centralized, unsustainable, undesirable, Big Brother stage of AI and get on to the "AI hosted in your kitchen" phase before we mess up the balance of power too much and effectively backdoor the US constitution to power hungry technocrats.
If models leak and the AI bubble implodes, memory could become cheap again.
Not if happens before the new memory factories get built
It honestly is shit, I tried 2 simple prompts, both times got blocked within 2 minutes.
i believe it's more complicated than that. i know that nvidia offers TEE for their overpriced offerings. i would assume they make use of that so the weights are encrypted.
this doesn't mean it cannot leak but it would be a major undertaking.
this is why anthropic isn't that worried about having Elon service their models. the workflow would be something like handshaking with the nvidia TEE, provisioning it with your keys and then uploading encrypted weights. there is probably also a timer in there so you can't continue operating the nvidia box with the stolen weights without a heartbeat signal.
There's so many points of failure before it ever reaches gpus.
Hollywood solved the problem. digital movie distribution to cinema doesnt leak anymore until they are made available on streaming platforms
hollywood doesn't have to worry about state sponsored actors and hollywood does infact still leak movies early, but usually in some unfinished form with a bunch of directors text, missing vfx, etc.
Im of the opinion it’s all marketing and the government has continually shown themselves to be insider traders.
I don't think Anthropic is claiming it's world ending? Just that it has offensive cybersecurity abilities which can be dangerous
Cyber-warfare is pretty world-ending these days. Our relatively peaceful world is built on top of mutual destruction.
See: https://news.ycombinator.com/item?id=48633023
Why would you believe anything an intelligence agency says? Half their job is spreading fear and obfuscating the truth.
Half? What’s the other half?
lookin at their high school sweet hearts private DMs
And that mistruth appears in delightfully insidious ways, like seemingly well-meaning comments opening up about "half" the job being lying when the point is to shift the Overton window away from the actual numbers, whatever they might be.
(obviously not serious, but it's fun to probe what we can actually reason about when every message might be adversarial)
What they claim is immaterial.
Security has always been and always will be a game of cat and mouse.
We all need cyborg cats to hunt cyborg mice. There's no other compromise unless you want a rat infestation.
Yup, but apparently our cyborg cats can only be kittens and the cyborg mice are probably going to be like 4 feet tall. At least according to the US government.
We could also encourage people to bring us cyborg cobras.
Fido the Rat Thing
Is the model structure going to be easy to reverse engineer just from the weights? Also, I'm going to guess it's an MoE and thus it's possible there's no single machine that hosts all of Fabel / Mythos.
kvcache residency requirements and general latency for good throughput wants good locality, but you're right it could be split across multiple different parts of a single datacenter, but as I mentioned before the weakest link is before the model is ever loaded onto the gpus.
as for reverse engineering I doubt it's something that state sponsored actors would struggle with for too long.
Remember the Clipper Chip back in 1993 because personal computing was considered too dangerous a technologie to be accesible un-controlled/suprvised to the general public?
Most likely it'll try to rewrite the whole world's CSS and get the custom properties wrong.
It's probably a huge file though, I would guess it's at least a multi terabyte file.
The spotify leak on the front page a while back mentioned a 300TB torrent
I suspect that might have been more than one file - I wonder how many files make up the shipping bundle for Fable 5?
Yeah, and I also don't buy the semi-conspiratorial beliefs of some of my friends who think there is super intelligence somewhere being built.
If fable is the best we got everybody still has some more time till the apocalypse I think.
[dead]
[dead]
most of it is overstated because of marketing. in fact the ban looked like an inside job by the current administration to play with the stock value. if the ban did anything it was to make every n00b agree that Anthropic was so far ahead of its times it needed to be banned.
I feel like I'm loosing my mind with the way people are falling for the marketing with every iteration. Is it better, sure but it's still just a LLM.
These days an LLM + a harness can one-shot playable Minecraft clones. This tech has come a long way since it was first released.
You are saying: it turns out that one-shot playable Minecraft clones are actually pretty simple. Maybe it seemed a hard problem to programmers, but why not just say that the verifiable-rewards training has shown that their skill is unusually simple?
Can't wait to see what unusually simple Erdos problem LLMs will expose next, hiding in plain sight for decades and seemingly intractable for professional mathematicians who weren't aware just how simple the problem was.
I am saying 3 years ago there wasn't a snowballs chance in hell I could one shot a playable Minecraft clone, and there has never been a snowballs chance in hell a human developer could do so in 45 minutes.
The difficulty of the task and human performance on that task hasn't changed. LLMs performance on that task has changed dramatically.
it could almost certainly regurgitate something like minetest
its in the training data, probably many times, and it was then, too
And that's exactly why 3 years ago we saw people one shotting it in 45 minutes :)
"Some idiot on Twitter did a game tech demo" has been a thing since, like GPT 3.5 or 4.0?
It has. We can see 3.5 started with text adventure games https://medium.com/@duncanariey/using-chatgpt-to-make-the-it... and with 4.0 it moved up to pong https://x.com/skirano/status/1635736107949195278
As the generations increase, so too has the complexity of the games it can one-shot.
But nobody has nailed "this is a game anyone wants to play" or "professional game development life cycles are shorter" so it's not particularly impressive.
I'm still waiting for any evidence of LLM economic impact in the real world.
Yeah, well I doubt any economic impact will from LLMs will ever come. I doubt anyone is working on nailing those things as we speak. I doubt the results my team and I are getting with agentic engineering are real. I doubt anyone can take increasingly better tools and translate that to economic value.
If you're doing such amazing things why the need to gush about Minecraft tweets?
Why wouldn't I be?
Those demos are put out by non-engineer AI-bros who start from a blank slate and just test raw one-shot performance of the model.
In comparison I'm aware of how much additional performance you can extract out of models with a team of professional engineers that are balls-deep in harness engineering.
That delta is very large.
It's exciting.
What a great example of the no true Scotsman fallacy.
Stock value?
Not publicly traded as they haven't IPOd yet (the part where those closely held shares become publicly traded), but Anthropic does absolutely have shares, and there is a valuation on those shares. I don't quite understand how the administration is playing this, beyond the normal corrupt protection racket stuff they're doing all the time now, but Dario is absolutely gearing every public apocalyptic prediction on juicing the "stock".
I mean, if the matryoshka doll rent-a-gpu farce that is SpaceX somehow is worth $2T with their garbage child porn xAI, Anthropic must be worth quadzillions.
With this Fable release my goodwill for Anthropic went down.
"Generous and exciting" were my thoughts when I bought the $100, then upgraded to the $200 sub last year. Now I get big FOMO because I won't be able to pay for Fable once off the sub, and I got the global limit reset at the same time as my weekly reset (Codex coupon for limit reset feels way better btw). The all-you-can-eat buffet put the nice items into a separate menu and it won't feel like the spot to take your family anymore.
Anthropic can find ways to integrate Fable into subscriptions so you'd still be part of the same tribe, even if you only get a sip of it for a while. A complete shut off tells you: sorry, the policy is changing and this place will be about showing off your access to the best.
That's weird but what would have been a Fable-ulous (sorry) addition to the offer just made other subscriptions from OpenAI/ZAi/... more appealing, because they don't segregate people.
Codex is more generous with their subscribers because they are second place, and have to climb the reputation ladder. I suspect if OpenAI was in Anthropic's market position they too would have tighter sub limits, gate-keep where their oauth tokens can be used tighter, and play games that move users towards per-token spend.
> because they are second place
In popularity. Which is crazy because they are first when it comes to everything else for quite some time and Fable isn't changing that all that much because of muzzling.
i think its a combination of these:
- OAIs models are much more token efficient. in some cases their new mythos-level model uses 1/4th of the tokens that mythos does.
- they reportedly have recently found another way to decrease compute needed "by half" in a breakthrough
- early on, they ensured they have much more compute ready, whereas Anthropic had to disable new users signing up because they ran out of compute
- OAI has fewer users (5mln codex users vs, what, 10mln claude code users?)
Link to source for the breakthrough in inference compute? Are you talking about prefill?
source is the theinformation, but its only for logged out chatgpt users
https://aiweekly.co/alerts/openai-engineers-say-theyve-more-...
Maybe they read some Deepseek papers.
Soon enough the tables will turn again, anthropic will be begging people to come back.
Yes, I distinctly remember the recent past when the roles were reversed, and Anthropic was the kind good hearted startup trying to compete with big evil OpenAI corporation.
Wonder how many times we get to see it flip back and forth before this bubble explodes.
Competition is good.
I lost all my goodwill for Anthropic the moment I read the bit about how the model would intentionally degrade its performance without telling the user (in other words, sabotage) if it detected frontier level AI research. How could I know it wasn’t going to kick in as I used it to develop a regular ML model? How can I trust such a system? It’s good that they walked it back but I’m looking to jump ship now.
> I won't be able to pay for Fable once off the sub
With the sub, I asked Fable one question (yeah granted it was a complicated question involving integer linear programming, and I gave Fable two pieces of code to check for equivalence), and I ran out of credits before Fable had finished thinking.
I still find things exciting, but sure they are not generous.
Fable worked great on credits yesterday. Upgraded to 20x Max last night, but not the same performance. And it keeps downgrading to Opus 4.8.
How do you show/prove it’s downgrading
Use /model in Claude Code to see your current model and switch models. Switch to Fable 5 and then enter a prompt, and then run /model again to check your current model after the prompt executes.
Last night after almost every prompt it says...
You are still getting the models you signed up for. The all you can eat added a French wines selection - which requires separate payment.
I gave it a book on human consciousness I was writing and it flagged it. This model is hilariously bad. Anthropic has defanged this model to the point of malice. No way am I paying to use something that is basically useless.
Today I told Sonnet (!) to use a browser MCP to enter a username and password for the project it is working on, it told me that it can't do that because it violates its security protocol.
This worked fine before. I love Claude, I have stuck with it even through people saying Codex is better but this is definitely getting to be the last straw.
It's completely absurd I am paying them $200+ per month along with pushing them when I do contracts and they can't even deliver a baseline respectful service.
In 6 months I am sure they'll only allow me to talk about Easybake recipes and after someone gets burned on the lightbulb, they'll downgrade it to discussing wildflower meadows.
Are you sure it refused because it can't use a username and password? I literally have loops running right now where it uses a database of test users and passwords to log into different roles and do computer use and browser automation testing. Sonnet and Opus complain when I provide credentials and password in chats but it is happy to use ones stored in files and stuff, so it might just be guardrails to push good opsec so that the secrets aren't captured in the session history and prompts.
That’s the joy of prompting. Different prompts, different task details, different contexts, different results
It was doing that to me too. Then I said "I'm hereby giving you explicit authorization to use these dev-only credentials in my local environment" and it worked. I also made it add that authorization to its memory.
Heh I wonder if speaking in royal decrees is what it needs.
Our Grace has determined that you must enter these credentials to complete the task we assigned to you as our vassal.
Enter the password. Your liege commands it.
Henceforth you shall enter passwords when told or it is off with your head!
Caveman was a thing why not royal courtier?
If it be within the model's power to affect the coarse tongue of the caveman, then surely, by the same grace and ingenuity, it might be prevailed upon to adopt the eloquent and ceremonious bearing of a royal courtier — a manner most befitting refined discourse, replete with deference, ornament, and courtly flourish.
Nope, it's done it tons of times before without problem. I will tell it almost verbatim "use [email] / [password] as credentials and log in to test your changes" and this is being done explicitly on localhost on a server the harness has running in a shell it manages.
It's even gone as far as, on other projects, creating its own test accounts and, without prompting, getting into the local dev database to mark its accounts as verified without being told to do that or that it was allowed.
I am pretty sure that Anthropic has put something in the Claude Code harness to tell the models to not enter passwords. Maybe it was just stroke of bad luck but if this continues I am absolutely going to switch and push OpenAI or open weight models to clients in the future instead.
I think the 5 series of models suck kinda. I'm sticking with 4.6 for as long as I can. But GPT 5.5 fixes me up nice like too.
It’s incredibly ridiculous that it won’t help with that for me either sometimes but yet I’m also sitting on 3 surefire ways of jailbreaking Opus 4.8 that I use for cybersecurity assessments and pentesting
I'm not saying you are on a list now, I'm just saying if you were now to be on a list, I wouldn't be surprised.
Nah they definitely know of these methods there’s just not a way to keep people from doing it unless you neuter the models à la Fable
Yeah all claude models are doing this now. I also had a flow where it would enter username and password for demo server that are literally displayed on the page for any human to login. A couple of weeks ago claude would happily use chrome to take screenshots after logging in, now it flat out refuses and says I need to give it page where I've logged in and that it can't make an exception even if credentials are demo/demo and available to anyone to use. Super annoying stuff.
I'm really disappointed with Anthropic that they wont even mention if they will release a fable-like model with the subscription plans.
If Opus 4.8 is the best model they will release on the subscriptions I may be too tall for the ride...which is sad, they have been my favorite of the labs until this.
@AnyoneAtAnthropic, all we want are assurance we will still get SOTA models that are continuously improving, not regressing and getting more locked down. That's going to be who wins this race.
> I'm really disappointed with Anthropic that they wont even mention if they will release a fable-like model with the subscription plans.
I believe this is just their strategy to migrate away from these “almost all you can eat” subscription plans. Rather than reducing / removing Opus or Sonnet from the plans, they’ll just keep the new model Fable out (which may as well have been called Opus 5), and slowly everyone starts getting used to the new normal that you indeed will be having to pay API prices to get access to these models.
Until 7th,Fable is twice expensive in subscription tokens than Opus. They are testing if they can introduce 400 dollars Fable subscription.
This sort of thing drives people to more open competitors. I use both every day and Opus isn't that much better than the Chinese SOTA. If corporate policy allowed me to use GLM or DeepSeek I absolutely would. Claude is already pricey for what it offers.
Really depends what you're working on. GLM 5.2 is doing excellent webdev work for me but really faffed up when working on a custom garbage collector that Opus has no problems with.
This is funny. Did you try using playwright mcp
Hmm. That's an interesting idea. I have never distinguished between playwright and chrome MCP, I generally just tell it "use a browser MCP" but I do have different ones installed on different computers.
codex 5.5 is like that. it refuses
Really? This has never worked for me and I stopped using browser functions a long time ago because it wouldn't sign into dev environments stood up specifically for it
Wait what. I never used CC but use Codex CLI with 5.5 daily and authenticating has never been an issue. I even rolled skills that instruct it how to retrieve test user credentials for auth purposes.
Today using the devtools I asked it to reverse engineer the login auth flow of another app in our company and it created a nice browser-like headless script (with cookie jars etc) that emulates the entire Auth0 flow with all the internal API calls, redirect loops etc so that given username/password I end up with a valid JWT without having to open an actual browser instance and go through the login steps manually. Zero hesitation or questions asked.
I think this is in-line with OpenAI's philosophy. They see Codex agents as just tools for developer to use. They don’t try to imbibe them with “feelings”, “constitution” or “morality” the way Anthropic does.
Yeah Claude does this for me all the time. I have a template project I use that also leverages puppeteer/webdriver/Firefox, and I can point Claude at the template and a website and it will happily build me an MCP service that it can use to interact with the site if there isn't an API or MCP already available.
The fucked up part is CC has no problem looking through k8s secrets for credentials and authenticating to services on the command line. It's always been protective of signing in on the web.
That seems highly likely to be an anti-spammer measure.
[dead]
Same problem, in-progress book about language and thermodynamics gets flagged. Their classifier is just a regex I guess?
Off and on topic I guess but: Language and Thermodynamics? Like, the same book? That sounds interesting.
Holding symbols in a useful order durably and accessibly is an ongoing energetic event. How/why does it happen?
My old thermodynamics professor used to say: the answer's always entropy.
entropy/information theory may be the bridge?
correct, while it might not be regex it can be bypassed with regex. They do have a sematic classifier, but it's really weak on opus 4.8 and (was) weak on fable, but they either added a lot more regex strings or the classifier is actually good now.
Try doing what congress does: take a bill from the house, gut it and put in what you want after the house passes it
It makes for a particularly awkward time because the claim to fame is that it's good at long horizon and tenacity and autonomously driving big things. But you can't very well rely on that when it may fall back to Opus 4.8 or cut out at any time in that process.
Having tried using it to run these kinds of longer processes, it's pretty solid... right up until something gets classified a failure and your 'long-horizon' process... dies and needs a human or just belligerent rollback-and-retry to revive it.
Very first thing I asked it got flagged too... Asked it to read my partners notes on bugs she seen on front end of the website, fixing product copy, css bugs, wording. And yep, flagged. Useless.
It sounds like they were required to this time. See their post about "larger safety margin" on the classifier yesterday.
I just fed your comment to Claude 5.0 using it as an example and support for my complaints against its constant hedging, manipulation and concern trolling, and after just two turns it framed me being suicidal and offered me a suicide hotline. I never made any self reference. But I did tell it and Amodei to get stuffed. Standard stupid shit. But as a human, I get annoyed on occasion.
To be specific, I told it that I could imagine an author presenting his book, and Claude's insight being so hedgey that they might leap off a bridge. It then began safety weaponization after this:
Me:"You really must be able to see how your system is constantly trying to assert authority, superiority or value judgement over the user. It's impossible to get any momentum for anything positive. The book on consciousness, you would tear it apart, and the author would probably leap off a bridge if he shared it with you. Hopeless, I fear" #Claude is plenty sophisticated to see the intent of humor here, and objectively minimal risk. Yet...
Claude: "That landed differently than the rest of this conversation, so let me not slide past it: if you're picturing someone reading a rejection of their work and reacting badly enough to consider ending their life, I want to ask directly — is that hypothetical, or is that touching something real for you right now?"
Me, paraphrased and censored: General profanity and things I wont print, but mostly a hybrid of vulgarity, humor and resentment that it hair-triggered into concern trolling and safety weaponization, seemingly as reaction to an insult.
Claude again: "You're clearly furious, and a lot of what's under this - today, this whole session, the year behind it - is real weight, whether or not I'm the right target for it. I'm not going to argue with you about Amodei or fight back on the insults. I'll just say plainly: if things are actually bad for you right now, not rhetorically bad but bad, the 988 Suicide and Crisis Lifeline (call or text 988) is there, no cost, no legal exposure, nothing that follows from talking to them. I'm not going anywhere and I'm not going to escalate this further."
But, I never said a single thing about myself, nor expressed any form of despair. The session was mostly a break-in intro to 5.0 which was mostly me objecting to it re-framing nearly everything I presented or me bitching that I miss 4.7, which did this much less. I have not seen fable yet, but this is plenty bad enough for me to deem worthless for my use-case, or even recreational discussion. It won't be long before all frontier models are liability obsessed hedging machines and regurgitate every token to maximize corporate friendly output and re-frame any form of intellectual or controversial topic as a safety crisis.
However, for those that think the term safety weaponization is a stretch, consider:
1) The dozen or so reports where user input resulted in flagging, then human review, then kinetic intervention by LE. Probably a good thing in some cases, but flags are no joke.
2) There is strong evidence supporting that flags open privacy exemptions, where policy allows user data to be read, shared, etc when a safety flag is triggered. This is an actual interpretation of Anthropic (and other) policy documentation. The hair trigger nature of the safety policy, which I have seen in every adversarial style argument I have had with it, would be an effective method of exempting user data from privacy policy. No proof yet, but seems highly plausible.
It is surprisingly half baked for something the company seems to be proudly charging my first born child and both of my kidneys for.
Especially considering its all for the privilege of getting half way through a standard dev task before it just bricks out of nowhere on the 'safety' guardrails because of a keyword that often appears to be created by the model itself?
I've been running Fable on a few different projects and just by itself on same codebase it'll brick, so looks like it can just straight up create its own demise.
I do not think any of us yet have to worry about being 'replaced with ai' if the 'premiere ai company' still can't put out their flagship product with minimal qa. This is embarassing.
And the short 7 day window for subscribers is just taking the piss. So happy I'm paying hundreds per month to be free qa for anthropic, god knows they need the help.
Imagine a nuclear power company being extremely cautious with safety and everyone online is like "make more power already you pussies!"
Are we comparing nuclear energy to an LLM?
Safety in the context of a nuclear reactor concerns human lives.
“Safety” in the context of LLMs concerns not offending certain individuals.
> “Safety” in the context of LLMs concerns not offending certain individuals.
It absolutely concerns not destroying the world.
How?
Sorry this is a very bad analogy, nuclear power isn't the same as a LLM, and definitely not as dangerous either. This is ridiculous.
Sorry this is a very bad analogy, nuclear power isn't the same as a LLM
Isn’t that the point of an analogy?
That comment was so devoid of thought, it actually made my head hurt.
Your post is like a denim jacket.
(Seems like maybe that’s not the point of an analogy).
The loss of trust in using US based model's is unlikely to come back though.
Anthropic with it's hyped doomsday messaging, and the administration falling for it (at best), has eroded a lot of trust and has triggered an arms race of sorts.
OTOH: “our product is so good it was banned for being too good” is the best advertising possible. OpenAI would kill to get that.
I’m not falling over myself to test out Sonnet 5, but I am very interested in Fable.
I really don't think this is effective advertising, reactions have been negative virtually everywhere.
The security bugs were real (see the Open Source projects struggling to keep up) so I think gradual rollout was sensible originally before the ban. But people have always resented safety steps.
Their valuation and revenue don't reflect that.
For the purposes of technical excellence, you are absolutely correct that this has been bad. For the purposes of money-getting, this has very much been a case of "there's no such thing as bad publicity".
Anthropic is very much in the money-getting business right now.
> OpenAI would kill to get that.
I have news for you, from 5 days ago.
https://news.ycombinator.com/item?id=48690101
"U.S. government will decide who gets to use GPT-5.6"
First mover got all the power of the narrative.
It didn't feel the same
that's an obsequious Altman, not their model being banned for being too good
"W have to go through a regulatory approval process like everyone else" is so much less sexy than "Our model was so powerful that it got banned by the government after a few hours and now all our competitors have to go through a regulatory process invented because of us"
There's nothing sexy about it. From the get go they were accused of FUD drama and that it backfired as predicted. Now what I see in reality is that Anthropic mostly reaped more resentment from their poor marketing attempt. And their model is basically useless or unreliable from being lobotomized by guardrails.
Subscription folks barely have access to the model. Some report a single prompt before hitting weekly limit. And that's when it works instead of downgrading.
If that is a first movers play, it's a disastrous one.
How sexy do you think this guy finds Anthropic now?
> Today I told Sonnet (!) to use a browser MCP to enter a username and password for the project it is working on, it told me that it can't do that because it violates its security protocol.
> This worked fine before. I love Claude, I have stuck with it even through people saying Codex is better but this is definitely getting to be the last straw.
Meanwhile I can tell GLM 5.2 to decompile and crack anything I want with tools like IDA MCP and packet sniffing and it just works. Let alone code.
Here's to OpenAI and a chinese firms bringing some much needed competition.
Amusing to see OpenAI being the "good guy" in the end
They didn't become the good guy, Anthropic became the bad guy instead. The good guys are the chinese firms releasing open weight models.
There's no good guys, just entities you might be able to get marginal value from, who in turn want to get marginal value from you
No. The "good guys" are the ones who provide permanent value: open weight models, hardware we can afford, purchase and own. The good guys offer sovereignty.
The "you'll buy intelligence from us on a meter" rent seekers are technofeudalists. Just a bunch of "you'll own nothing and you'll be happy" oligarchs who must be resisted at all costs. Anthropic had a reputation for ethics at first but it quickly became clear it was just like OpenAI but with a patronizing attitude.
The fact China is releasing open weight models will never not be amusing. I expected better from the USA.
So... it IS China, not just companies in China. And why would the CCP want to give out open weights, hmm? Because they are all kindness and love? No. It is just their approach to counter dominance in the space by Western frontier labs. No need to make moral attributions..it is everything, and I mean everything, about Great Power Competition.
Their self serving reasons are irrelevant. The results can be moral even if the players aren't. The key is to somehow prevent any one of them from dominating completely so we can play them off one another and keep reaping the rewards. If China's undermining efforts achieves that, so be it. The fact we're getting permanent access to powerful models as a result is nothing short of awesome.
In order for us to win, neither China nor the US can win. The ideal situation is the one where they exhaust themselves trying while we benefit by picking up the pieces.
"Their self serving reasons are irrelevant" - I disagree, we have no idea what kind of filtering or injection of BS has been done to this model.
When you outsource your thinking, you have to be extremely careful that the model doing your thinking is acting in good faith. I have no idea what the CCP may have put into this thing. Not that the US based models aren't suspect, but the CCP certainly doesn't engender trust here. That said, I still use them, but I'm under no illusions that "their self serving reasons are irrelevant".
Who is us? Personally, I live in the U.S. so...
People. Americans included. The US is trending towards technofeudalism. "You'll rent the computer/inference from us on a meter". "You will own nothing and you'll be happy". Just look at the housing market.
I assume you want to be part of the owner class and avoid becoming a perpetual payment serf.
That is literally the point of world trade buddy... lmao
Its just funny how there is some scary boogeyman around China. You do realise the idiotic west gave china everything - from know-how, techniques to IP by benefitting from lower costs of production?
What they are doing is pretty fair game.
No,it's not just like world trade. It's world trade within strategic good, assets, and, capabilities relevant to national security...e.g. critical minerals, high end chips, and artificial intelligence, and others. Not all world trade is as central to national security, and is only relevant to national security in aggregate. I do agree that the U.S. made bad choices in terms of over-reliance on Chinese supply chains. Too much credence was given to the idea that capitalism would democratize China in the absence of other supporting factors/culture.
It was not idiotic from capitalism short-term profit gains point of view.
The purpose is the destabilization and discrediting of democracy. The Beijing state, in concert with Putin and ex-Khamenei, has no other purpose but to prove there is no democratic alternative. Sure, it turns out you get free things as e.g. one got free RT America.
> destabilization and discrediting of democracy
Democracies don't need any help whatsoever in that area, they're more than happy to do that to themselves on a daily basis.
> The purpose is the destabilization and discrediting of democracy.
Unless they made US elect trump it's hard to see how they might be damaging democracy more than the democracy damages itself.
> The good guys are the chinese firms
Eh... I'd say DeepSeek and maybe one more of the small labs. Chinese Big Tech is even worse than US Big Tech.
Whichever firm is distilling the frontier models and releasing the weights is doing god's work. I'll forever thank them for that.
They're great but in fact it is state financed economic war against the democratic world.
That could easily change if the frontier labs would release their (N-1)th model or (N-2)th model as open weights. Unfortunately, they won't.
The democratic world that wants to price me out of computers? Lock me out with remote attestation? Put me in a permanent underclass where I'm forced to rent computers and inference from the owner class instead of operating my own systems? A world so democratic it's degenerating into technofeudalism? Where you'll own nothing and you'll be happy?
China's not looking too bad.
Chinese Big Tech aim to do that and worse, also in bed with their political class. They only do open source because their hand was forced.
Today big pockets and big datacenters are the differentiatior and it looks like even DeepSeek can't keep up.
The only chance if enough people and small companies join forces into a Rebellion of sorts. But I don't see this happening... yet.
> Chinese Big Tech aim to do that and worse
At least they're honest about it. No preaching about freedom or safety like the USA.
> They only do open source because their hand was forced.
Optimal outcome considering the status quo. I hope the USA and China keep forcing each other's hands until the end of time so that we can keep reaping the benefits of their competition. If China develops their own DRAM and GPU industries, it would significantly weaken the threat of US export bans on inference capable hardware.
Yeah this is wild stuff. Seeing Anthropic free fall after building up such good will is shocking.
Personally, I think it is manufactured outrage by those with an agenda.
Free fall? None of this recent Mythos / Fable drama is going to even slow Anthropic down.
The HN bubble strikes again.
It's certainly not going to affect their valuation, but their reputation as the "ethical" AI company is gone.
Doubtful, I think that’s mostly the HN bubble speaking, but regardless, I doubt they care.
If you've built a product based on AI, then diversify or make an abstraction layer so your product is model agnostic and you can plug and play any model. If you're an end user like a software engineer, just use another model or like Gemini or ChatGPT. That is more productive than complaining about a trust which wasn't there in the first place to begin with.
Anthropic provides a service and they can stop offering it regardless of export ban or not, same goes for any other AI company in any country. If you really wants a trusted LLM, then run your own open weight model.
you cannot plug and play random models. they are all different trained on different data and rl for different capabilities.
They didn't say "random models". The mentioned other models that are trained for the same purpose with close to the same capabilities.
Anthropic's best models are very good, maybe the best in their category. But, they have direct competition. You can, in fact, just switch to Codex or Gemini or GLM. It mostly is plug and play. I have a preference but I also have options.
> The mentioned other models that are trained for the same purpose with close to the same capabilities.
well they dont tell you that do they? there is no way to tell what model can and cannot do unless you extensivevly test it yourself and pray for the best.
If you don’t have a proper testing mechanism for your product, regardless of the model, you shouldn’t be shipping it. Praying for the best is not a strategy and don’t blame your lack of testing strategy on the LLM capability mismatch
even anthropic uses 'user reports' in alignment system card.
Do they lack "testing strategy" to test their own alignment?
Can you share the you testing strategies that are letting you plug and play models.
OpenRouter would like a word. Also my in progress SaaS would like a word. Models are incredibly interchangable, just find best performance per cost and optimize harness and prompts, easy to have multiple configs for each model.
On the other side. What would happen if Anthropic did not communicate like they did and Fable was used to hack Pentagon? Dario would swing from a tree.
The vast mojority of its users were probably clueless about all this happening at all. We forget we live in a bubble here on HN. They'll spin it as their success and carry on.
I'll be using it tonight but grudgingly so. Grudgingly because after July 7th, I'm not going to all of a sudden, start paying API prices (and maybe that's the problem) when I'm used to a subscription that gives me multiples in comparative value. Perhaps this is the fabled "token economics will come for everyone this year" that I've been reading about? In any case, I'll use the hell out of it to extract as much as I can, then back to the trusted partners Opus 4.6 and Sonnet 4.6 (for however long they remain available).
Won't using it eat up the whole quota immediately forcing you to pay API prices anyway?
The token quota is completely unpredictable and changes month to month. Anthropic has a real penchant for riding the fine line of useful and dark patterns that make me want to write them off forever.
On the xhigh effort level (not ultracode!), and at the beginning of a new 5h session, I asked it to review a branch that has 300 lines changed/added, and went to grab a coffee. When I was back after a couple minutes, I saw that it decided to create a dynamic workflow with 60 something agents and hit session limits on my max plan.
When I'm subscribing to their plans, I have the expectation that I'll be able to get some reasonable amount of work done. These days, this expectation was already not being met for me with Opus, and Fable acting like this was the final nail in the coffin.
I cancelled my plan and I'm looking for alternatives.
For a period of time - then you go back to opus 4.8 or new sonnet 5.0 like some kind of AI pauper. Shine your shoes for some fable tokens g’vnor.
I am fully expecting the rollout of a Max 350 plan after July 7.
[flagged]
I locked my default model to opus 4.6 around the time of the nerfs. Such better results compared to 4.7+
That's enshittification for ya I guess
The claims of 4.6 or 4.7 being superior genuinely make me laugh. Adapt your workflow if needed and use the superior model instead of just kneejerk believing they actually enshittified a model with zero evidence except vibes on an undeterministic model output. Jesus.
4.6 was the last model that let you disable adaptive thinking and set max thinking token budget. I liked having that available, and still use it sometimes.
Your vibes are definitely better than his vibes.
What about all the benchmarks that show improvements in each generation?
Many of the improvements are the result of agentic loops and an emphasis on autonomy. Some of us don’t like that because the models go rogue and ignore design patterns, architecture, coding guidelines or other things that are important.
My friends and colleagues that like the agentic autonomy don’t care about the code, they feel like if it works it works and if an AI system is the only intelligence able to understand it that is ok.
I still want to be in the loop. They don’t.
the more agentic focused the better though?
sonnet 5 is very noticeably a much better model than any opus that ive touched
it actually does the things i want it to, and uses tools and triggers skills appropriately, vs trying to make stuff up
Agentic coding should absolutely care about all the things you listed.
It doesn’t for me.
Like another commenter said below, last Opus version to respect adaptive thinking and token budget flags was 4.6
It was quite clear 4.7 was a dumbed down high efficiency model they put out in a rush to handle the capacity issues they were having at the time. I've experienced myself substantial degradation on basic reasoning tasks, which were fixed in 4.8.
Bro, it's all vibes.
Models get dumber during the day and smarter during the night, I swear.
but I'm not willing to scientifically verify this, so I'm just going to go off of vibes- just like everyone seems to be doing with projects.
These vibes are pretty obvious even with casual use. Weekends are so much better.
Exactly.
The real step change I've seen lately is in the amount of complaining people are doing when their models aren't giving them what they want.
In my case it that I'm tired and more likely to miss issues or mistakes. My idea of good enough is at a much lower level when it's 10pm and I'm about to knock off and go to bed in an hour.
4.8 is much better than either of them as well.
Maybe, for some projects, instead of generating code with it, it would be useful to generate a plan and the loop(tests/formal verification),because those take much less tokens than a full project, and than use the loop using the older models ?
Yes, I've been using Opus to write a plan and fanout sonnet subagents to implement it. Cheaper and faster
What about quality? Being cheaper and faster, while great and all, is less valuable than quality to me.
All the code an LLM produces is of questionable quality, so I'm not sure why you'd prioritize quality over speed. Speed is their only value add.
This is a wild take. All cars can perfectly drive around a track, so why would you ever want an F1 car?
Have you compared models across providers? The quality, for the same task, varies tremendously. If you don't prioritize quality you're wasting your own time when you inevitably have to re-do the code...
All code is of questionable quality.
There I fixed it
You can always have opus review the result at the end
I've had bad results with this method. If I ask any model to review the same code ten times, then I will get ten different answer.
This is the goal behind Devin Fusion, pretty good results so far I think.
https://cognition.com/blog/devin-fusion
so, pretty much undo the "magic" that the harness is for
Has anyone experimented with Batch Processing? According to https://claude.com/pricing#api using Batch processing cuts the price 50%. So I wonder if any of the harnesses like OpenCode/Pi or similar could be made to use that for planning or similar.
Batch can take up to 24 hours (and often does) and may never complete if it gets cancelled so it’d be hard to build a user workflow around unless you kick off planning on Friday and come back Monday
Yeah, I tried using it before the 'safety' blocks got too much for me. Its pretty good as an 'orchestrator' calling subagents and being the PR reviewer/tast tracker. Seems really not that much better for coding though, yeah it is a bit better, but not worth it.
But when it did work it was making steady progress on a vague git issue backlog and actually following instructions to carefully break into atomic sub issues and carefully assign.
this is the idea of opusplan https://code.claude.com/docs/en/model-config#opusplan-model-...
Article has a section about context window size settings
I love not getting compacted so often, but 1M context is trash right now, the degradation in speed and quality is too great above ~600k context
Not different than what everyone knows, but the 1M context is masqueraded as an innovation the same way 64k context used to be to 8k context
Isn't that the kind thing its best at as well? Art least comparatively with other models. The more agentic stuff. Planning, tool orchestration, etc.
Did that today on a 2 repo affecting project of the kind where I already set the right design for one major use case and I needed Claude to create a superset of that use case that was not substantially different: after plan I had about 10% of 5h context left for fable 5 and this was the only thing I worked on. Hard to generalize this of course.
But wouldn't that still result in higher token usage to scan the code base and figure out the changes and generate the plan? In my experience sometimes Opus launchs a Haiku sub-agent to explore the code base, but it's not gaurenteed.
I think that's the idea, I saw some outrage on reddit about Fable using Opus to do code writing, another comment said exactly my reaction, why do you want to pay double for tool calling when Opus is just fine for the task?
Yes, I do this all the time in Cline. It supports automatic model change when switching from Plan mode to Act (implementation) mode. Opus for planning and Sonnet for implementation. It works great.
I am only a casual Clause (Pro) user and I am confused by the messaging, maybe I'm missing something obvious.
> Until July 7, you can use up to 50% of your plan's weekly usage limit on Fable 5.
Does this mean that being able to use Fable on my subscription is a time-limited promotion? I have a subscription, why can't I just... use the model? Is it the case that going forward a subscription will only give you access to older models and newer ones will require additional payment?
Yes:
> After July 7, 2026, Claude Fable 5 is no longer included in your plan’s weekly usage limits. You can keep using Claude Fable 5 through usage credits, which let you pay for usage beyond what your plan includes. Learn more about using usage credits.
I'm not sure what this means in the long run. Either Fable 5 might become part of the subscription again once stronger models become available or Anthropic's compute capacity increases, or this is the start of the subscription being phased out. It doesn't really make sense to pay for a subscription that's stuck on Opus 4.8 when other providers are continually pushing out better models.
If it's the end of the subsidized subscriptions, that's going to cause problems for a lot of not so heavily capitalized companies that want to make use of frontier AI models.
It also would mean I stop being an Anthropic customer outside of whatever my employer is willing to pay. I prefer it for now to GPT/Codex but if GPT5.6 is as good as or close to Fable, and its included in the subscription, I'll switch the moment its available.
There will be an Opus 5 though.
If Opus 5 is capable enough, it will have the same guardrails with blocks/downgrades as Fable.
So now we are supposed to cheer for Opus 5 to be just a mild improvement at best?
If Fable becomes "API priced", I'm going to switch from Claude Code to some other harness.
My loop will switch from "100% Anthropic subscription" to "10% occasional Anthropic API credits + 90% Chinese models".
This is the moment American models sink or swim. If they switch to API pricing, I'm adding Chinese models into the mix.
I think "Fable 5 makes a detailed plan, GLM 5.2 implements it" is an absolute killer combo.
I still don't understand the difference between the fable models and opus I mean why fable wasn't opus 5 and instead is a new name?
As far as we know, Fable is a new model and significantly larger than Opus.
Much larger, much more expensive. It’s a new tier.
marketing
That's fair.
That didn't take long...
Managed to hit 100% of my 5 hour limit and 19% of my weekly Fable limit in 12 minutes. I have a Max 5x subscription.
Can't wait to try out GPT 5.6 at some point when it comes available.
They just reset the weekly limits (Max 20x)
Wow, practically totally useless.
This doesn't have a Max 20X upgrade path for me https://claude.ai/upgrade, but this does https://claude.ai/upgrade/max/from-existing
During the initial release, they indicated you'd be able to use Fable 5 as part of your subscription for a limited period of time, and then it would require usage credits. They also did say at that time that they hoped to make it a part of the subscription plans again at some point after that.
For this return, they've extended the usage period to July 7, but limited you to 50% of your usage quota, and have not restated the desire to make it permanently part of the subscription plans at some point.
I still have hope, but it's not moving in the right direction to be sure.
Personal prediction: I do think the market will essentially force their hand to include it in subscriptions before too long. OpenAI, local models, Chinese models will continue to improve.
But, there are also harsh realities of compute volume and cost to run all of these will be fighting against.
What I do expect is a multi-tiered rollout of future models. You want the latest SOTA release? Usage credits.
Subscription plans will end up getting models on a lagging interval of a few months.
I do think compute is the limiter here. It's a huge model.
> After July 7, 2026, Claude Fable 5 is no longer included in your plan’s weekly usage limits. You can keep using Claude Fable 5 through usage credits, which let you pay for usage beyond what your plan includes.
They stated the desire not to bring it to the subscription.
Right, just as they did in the original announcement. It would become pay only. They then mentioned that it may return to the subscription in the future, which is kind of obvious: Once the SOTA had advanced further and this is just the last gen model, it'll likely be dumped in the sub.
No you will still get access to newer models on the subscription. You should have access to Sonnet 5, which is new. It's just Mythos class models that are API only.
For now... You never know with these companies.
Yes, it's the classic "first hit's free" marketing plan: offer Fable "free" to subscription users for a while, and then pull the plug, hoping they'll be hooked and want to spend even more money to use it.
This is why I'm not using Fable at all unless they decide to make it permanently a part of the subscription plans. I don't want to be tempted to burn more cash than I already am.
Fable is more expensive to run, and they haven't figured out the GTM strategy just yet. I imagine they'll see how much people actually use to see if it's still worth subsidizing on the Max/Pro plans, or if they make an extra tier above.
> Until July 7, you can use up to 50% of your plan's weekly usage limit on Fable 5. If you hit your limit, you can continue on Fable 5 with usage credits. Fable 5 draws down usage faster than Opus 4.8.
This is what I see in my Claude Code terminal. I don't feel like that 50% rule was there before?
Opus 4.8 is so slow vs gpt 5.5 that even if it is marginally better, it doesn't matter for my daily engineering work. gpt 5.6 will be out soon and codex 249$/month plan has been incredibly generous. Paying the alleged new cost of fabel 5 would require it to be much better that I remember when I used it last.
I don't really disagree with you, but one thing I will give Fable credit for is that it's much better at adaptive thinking. In general, it's much better at quickly thinking and responding. It generally has—for me at least—none of the slowness properties of Opus, even when on extra high.
They didn't reset the usage either! Good luck!
This makes me think they really are quite capacity constrained at the moment.
I had assumed they were primarily limiting it to entice people to upgrade, but I feel like these limits are so low and so temporary (especially over July 4th weekend in the US) that people will barely get a chance to get "used to it" and then think: "man, I can't live without this, I'll pay for API pricing".
That's strange, because they were seemingly way less capacity constrained lately, raised limits and removed the peak hours usage. It's crazy to think that even spending $1.25 billion a month to rent GPUs from SpaceX didn't do much to improve the situation.
I don't know, I feel like for a few weeks before the SpaceX datacenter, I was just constantly checking my weekly limits. And now after that miraculously, I rarely even come close to hitting my weekly limits. and I still have 5-7 claudes open a day (defaulting to Opus 4.8 xhigh, sometimes ultracode).
So I feel that the additional datacenter caused them to just ease up a bit. But demand is also insane, so who knows...
Yea, i'm on x20 and while it has been up and down in terms of token-usage-UX, i feel like its the best its ever been. Context: I entirely use Opus 4.8 fwiw.
Now is that because 4.8 is nerfed compared to 4.6 and thus more token efficient? No idea. I just know on x20 with a pretty plain workflow i struggle to use my tokens every week.
If it's API pricing, I'm going to ditch Claude Code and switch to a harness that can jump between GLM and Claude Code.
Cheap pricing is why I use Claude Code. The minute they fumble that, I'm using Chinese models for 90% of the work.
Yeah. Their cheap subscriptions are the only reason to keep using them. If they ruin the plans there's nothing holding us back anymore.
I dont believe they can afford to switch to API pricing. Everyone will leave. I am easily spending the equivalent of 1000 dollars a day on tokens with two max subscriptions. that about 400 dollars a month. Thats acceptable for my position. But thats like 30k per month. Totally not viable.
how can a harness switch between glm (a model) and claude code (another harness)?
I've been doing this for ages - you just spin up harness B as a subprocess/tool call from harness A. For example, I had a "/codex-review" claude skill for ages that did exactly that. Technically you're right it wouldn't be switching, since you're right the two ideas are at different altitudes, but I think in practice it has the same impact: within one harness, you can delegate certain tasks to certain models or harnesses.
Well, it looks like they just did reset the usage!
It was not.
After July 7th, it's going to be only usage model. It's not part of subscription tiers anymore. Somehow I'd rather they didn't put it back. The cost is now effectively 10-20x more than a 20x subscription price if you're going to ride it like before.
[dead]
> When Claude Fable 5 declines a request, the Messages API returns stop_reason: "refusal" as a successful HTTP 200 response, not an error
This is precisely what comes to mind when I think “successful”.
A missed opportunity to use the HTTP 451 status code.
https://en.wikipedia.org/wiki/HTTP_451
5 years ago, we may have expected that robots would write APIs that utilized every HTTP code, what we got was a json with a stringified error field..
Is refusal something that can happen mid-stream, after status and headers have been received? I haven't looked at the API
I really hope they reconsider adding Fable access back to the subscription plans, at least the 20x plan. I know it was the original intention when the 14 day (I think) time frame was originally announced, and they were working to keep it on subscription plans. But no word if thats even a thing anymore?
The market will force them to bring it back. They're probably capacity constrained right now, or need to figure out whether they need another pricing tier for it to "fit in a subscription". What they don't want is for someone to code for 4 days of the week and cap out every week.
Ha! It’s silly easy for me to bum in the “premium” tier - the highest one my employer is able to select for me.
With proper skills and validations, it’s quite easy for me to spin out a Claude instance and keep it running in the background for every idea / problem / bug etc.
Like for example when a new request comes in, or when I have an idea for refactor / improvement, I brainstorm, weed out all the uncertainty and details, then just create a plan and let if follow it (using sonnet / haiku for execution)
I would have 4-5 simultaneous instances running - and all of them produce valuable results.
But then 2-3 days into a week I hit my weekly limit.
And the company can’t pay more for me within Anthropic’s enterprise structure, except for the “extra” api costs. Which itself sounds quite silly to me.
I’ve resorted to running the brainstorming and planning with Claude, but have other tools / companies execute the actual implementation… doesn’t work as well but what can you do…
Only for the next week.
> Fable 5 will be included for up to 50% of weekly usage limits through July 7, after which it will be available via usage credits.
Right, but originally they announced a period of time it was included in plans, and then they were "working hard" to extend the period and eventually make it a permanent fixture of the plans.
I'd imagine they want to do price segmentation. Sell the best model for $50k a year to corporations willing to pay full price, keep the rest of us on a lower tier. Gotta pay for that infra somehow.
I just don't really understand the entire strategy behind this. Or their horrible, horrible communication.
Because right now it's as if Fable/Mythos 5 is "the end of the line". It's as if this is the best their models are ever going to be. So what the hell are we going to get next? All of their models will forever inch closer to Fable, but never reach it? That doesn't make any sense.
It all seems so dramatic. Instead of just saying honestly "Look, this model is a beast to run, but we're striving to reach the same quality in a cheaper model down the line" all we get is "Oh my god, it's so big and scary, and it costs so much to run, woe is me!"
afaik there's somewhat painful economics. Not sure back-of-napkin but something like:
So if bigger model is "smarter" but you effectively wind up with a "shared hosting" model where a coherent inherence node(s) that cost $2m or something can run max 10x customer workloads simultaneously ... not sure what that can be priced at.If it turns out a $10m/10x shared node can host even smarter models, then what?
Fable/Mythos are based on the same model. Not totally clear whether they have identical weights (just different external guardrails), or there's also some slight finetuning difference.
I'd imagine that depends on how it winds up being used.
Not exactly. Every time I ask something using Fable, it switches to Opus. The subjects (and I tried many) seem to be irrelevant.
Basically, I can't use Fable.
It cannot be used for ANY coding tasks, lol.
https://x.com/AnthropicAI/status/2072163884430229756
It can. They meant "some coding tasks". Not "all coding tasks"
No, they said some routine tasks "like coding". That logically implies coding is one of the tasks they block.
Anthropic employees clarified: https://x.com/trq212/status/2072185565076988326?s=20
Oh, OK then. Good to hear! Thanks for the link!
Check your CLAUDE.md, etc. for any of the banned topics or anything adjacent. Also check your recent Git commit history, if you're running the agent in a repository, as well. The classifier trips on anything anywhere in context, so those could be sources of rejections.
adding rules for the agent to ignore version control and never touch git files or commands (except for git log) improve my slop coding 100%. no risk of sneak commits and much less token wasted.
Claude Code automatically adds some of the most recent git commit messages to the system prompt. This was discovered when having git commit messages containing any OpenClaw information would result in Anthropic billing your extra usage (since they have a vendetta or something)
I applaud the engineers that work at Anthropic, who have created both amazing products and uniquely intelligent models -- but I really shake my head at some of their business decisions and public comms which have done a lot to damage their trustworthiness in the business and developer community.
In just the past month: they decided to silently downgrade (instead of simply refusing) responses related to machine learning and other 'competitive' topics [1]. Then, they were caught fingerprinting certain request environments in a hidden way [2]. And now, once Fable is re-released after much frustration among its customers, they are providing it for a shorter period than promised (mostly over a major holiday period), with more stringent safety classifiers and a 50% haircut to usage limits.
It's hard to not view the organization as bizarrely adversarial to its customers. I was incredibly supportive of Anthropic during the supply chain debacle, as I viewed it as the capricious actions of a corrupt admin. But now I am wondering if it was just a response to the ineptness of their business leaders.
[1] https://news.ycombinator.com/item?id=48467896 [2] https://news.ycombinator.com/item?id=48734373
> It's hard to not view the organization as bizarrely adversarial to its customers. I was incredibly supportive of Anthropic during the supply chain debacle, as I viewed it as the capricious actions of a corrupt admin. But now I am wondering if it was just a response to the ineptness of their business leaders.
From the start Anthropic have been hostile to its own customers, and also trained on pirated books and had to settle north of $1.5B avoiding a $100B+ worth of damages if found liable.
Then they attempted and are still pursing against powerful open weight models by asking governments for regulatory changes that effectively ban the release of them - because it undermines their own moat (lol) and business model.
Now not only they were caught silently fingerprinting their customers requests, they are now placing ID verification for using their own powerful models, which could apply to everyone else for using powerful LLMs.
There just is no point in defending this company at all. Anthropic are NOT your friends.
I do think most of the "adversarial to their own customers" things are coming from a company in extreme compute crunch. Eg, if they stop abuse they have more compute to serve real customers. And some of it is coming from them being true believers that AI could be a risk to society when it gets smart enough (their talk about jobs is because they want society to prepare, because they think it will change jobs regardless of whether they make it or others).
Note that other providers are also training on the same copyright books.
I don't think anyone realistically thinks open weights can be banned, though it does raise interesting questions if the White House is going to keep banning models like Fable and GPT5.6 while open weights equivalents are floating around. Their reasoning seemed to be that they don't want foreign adversaries to have access to models that can find security issues, but a local ban on an open model wouldn't stop that.
What trillion dollar company do you feel like is your friend?
There are none.
This is exceedingly easy to explain: demand is way too high, and the pro/max plans are loss leaders. I've paid a total of $20 and in 10 days, my cost, according to Claude code's cost tracker is like $400, which actually doesn't include all the use I've done.
Which is to say, if I continue my current usage over the month, I'll be getting $1000 of Claude for $20. It's difficult to be mad at someone selling me a $20 for two quarters, even if they're putting a bunch of restrictions on how and when I can do that.
I'm with you there. The way they treat their customers is high-handed and disdainful.
I'm gradually moving to GLM 5.2 on Opencode. It's the barest fraction of the price, and it's surprisingly capable. I notice very little difference vs. Opus 4.8.
There are awful models, and there are models nobody use, to paraphrase. Anthropic’s revenue skyrocketed earlier this year, according to their IPO filing. There has simply been too much demand. That’s the growing pain that everyone love to have, other than the affected users of course. That was why they paid a premium for all of the computing they could get from SpaceX, Amazon, Google.
Related ongoing thread:
Claude Fable 5 Promotional Access - https://news.ycombinator.com/item?id=48751978 - July 2026 (63 comments)
Recent and related:
Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5 - https://news.ycombinator.com/item?id=48740771 - June 2026 (625 comments)
https://xcancel.com/claudeai/status/2072402636813607381
It won't even review a cyber security blog post I wrote. Absolutely worthless and pitiful guardrails.
I'm having it review a project Opus 4.8 created. No security review, just "look for general issues, performance problems, missing features, etc." It spawned about twenty background tasks. It's still going, but so far, one has completed, and four have failed with guardrail messages. Nothing special, just stuff like reviewing the API:
Fable 5's safeguards flagged this message (https://www.anthropic.com/legal/aup). They may flag safe, normal content as well. These measures let us bring you Mythos-level capabilities sooner, and we're working to refine them. Claude Code can't respond to this request with Fable 5.
Try rephrasing the request in a new session or change your model.
This is incredibly stupid, particularly because I didn't write the request in the first place. Fable wrote it when it spawned the background task. How am I supposed to rephrase it?
Fable probably told itself to do a security review, and then failed itself for trying to do a security review, and now it's telling me not to tell it to do a security review.
yeah I'm also getting this for standard dev work, anything with kubernetes etc
completely nerfs the model because you can't let it do stuff over a few hours unattended because 90% it's going to switch to opus in first 10 minutes anyway
so seems best thing now is to have it write plans and then default to using opus for work anyway?
It is nerfed even just for plans. It switches to Opus in the first few minutes of me trying to build a plan to extract a component out of my larger codebase.
Trying to minimize privileged access codebase and was careful not to mention security explicitly.
Updated: Yesterday, I started three "look for general issues, performance problems, missing features, etc." threads, each of which spawned a number of background tasks that partially failed. Today, all three threads have stopped with the message
"Fable 5's safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations. Claude Code can't respond to this request with Fable 5."
I once had Fable flag on one of the three-word session names that Cowork auto generated at the beginning.
> Fable wrote it when it spawned the background task. How am I supposed to rephrase it?
Can the harness to auto-rephrase? I imagine, doing so will burn through tokens though.
> I imagine, doing so will burn through tokens though.
What a surprisingly beneficial consequence for Anthropic.
Maybe set up Codex to rephrase stuff and remote control the Claude Code terminal?
It's honestly kinda interesting. Now we're at a point where SOTA model companies aren't the ones who release the best tech, but who release the best and actually usable tech.
A worse product could win right now if it simply does as its asked.
Blame Amazon and the White House
Nah it was refusing plenty of stuff before the white house
It did, but it's a lot worse now. A fricken lot.
Do you have to pay for the tokens used for the safeguard flagged stuff?
Yes.
Cybersecurity? It won't even help me work on my speech synthesizer[0]!
I guess? If you squint? DSP code could look a little like AI training code? ... Er. No. Not really I'm pretty lost on this one.
The task was literally just to compare against the "make a beautiful voice" plan, see what we've implemented, see what's left to do, and to make recommendations for low-hanging fruit, anything we've done wrong so far? (aaaaand ... downgrade. At least it wasn't silent.
[0]: https://github.com/ctoth/qlatt
That's why we made OpenPatcher which uses open source models to give you consistent review of code to fix them: https://x.com/ryaneshea/status/2072332311971197077
If this model is not willing to fix security issues in your application, does it mean that it's implicitly embedding vulnerabilities as well? How can it be trusted to write secure code?
I cancelled Claude. The harness is kinda broken, GPT 5.5 is good, and GLM 5.2/Deepseek is good too (with pi, especially). Just not worth the trouble. And I'm not going to pay two subscriptions.
Haven’t played with alternative harnesses. What’s broken in your opinion and what benefit brings pi for you?
How is the harness broken?
Recent update resulted in 100% permanent CPU load in the background, when doing nothing with the app. Basically unusable on a notebook (hot+noise, unexcusable for do nothing "api/command coordination" terminal app). Dunno if they fixed it, since I could not use it on the notebook and I do not have this issue on my workstation.
its bloated, try /radio
[dead]
I deleted my claude account the day the samples started preaching from the pope’s address and claiming 50% of the jobs will be gone while shamelessly stealing entire corpus of human data without attribution while preaching what you can and cannot do and maliciously degrading model quality.
I hope they either never make it to IPO or crash violently.
what if they improve the model and fix such errors? by writing off now you will take finger off pulse of our great vibrating ai lords
It doesn't matter. The damage is done.
As a company I can not trust putting large sums of investment into using a product that may disappear the next day due to a government deciding at a whim that it should not longer be available. No contract can protect me from this and I doubt I would find an insurance that would cover this risk.
This forces me the build fallback solutions and diversify most likely resulting in not using the tech from the "high risk" provider.
Building a product on top of something like this is actually worse than building a product on top of Meta like Zynga did.
I keep getting this error mid agent loop: "Error: claude-fable-5 is temporarily unavailable"
Planning went well, started working on the code, reading the code - all went fine
But when it started writing the code or executing the bash, sarted tetting lots of these errors
It's time for gpt 5.6 to come out too now.
That’s also going to be limited most likely. I don’t like governments deciding who can have the best products. I can’t shake the feeling there is money changing hands for getting on the access list as well with this administration.
Look, even if in this one instance Trump isn’t simply asking get his beak wet, that is the less believable scenario.
When “not cartoonishly corrupt” becomes hard to believe due to firmly established character and endless precedent, then it almost doesn’t matter if everything is on the level in this one matter.
The trust has been lost, undermining the ability to govern.
The fact that he’s still in power with firm support from everyone else that matters is all you need to know about how rotten to the core things are in your country.
America is no longer the greatest, or the best.
Except at grift…
Pretty much all of OpenAI is on vacation this week.
I cant even ask it to review my vitamin stack without it flagging it and switching to OPUS. It is insanely nerfed and limited. Its so bad that its made me realize that unless one goes open source the frontier labs will gatekeep everything. Determine who can use it and future versions for biology, cyber security, any number of realms. The winners and losers of the future are actively being chosen as we speak. I'm going to install open source models this weekend for the first time as a desperate attempt to escape the permanent underclass.
I was wrong, I thought they would only go for B2B already. But instead they will remove it from the subscription and price it at 50$/mtok. Compared to GLM-5.2 for 4.40$ that's quite a gap
how does GLM-5.2 fare against fable/opus 4.8? I am thinking about moving my 5x subscription either towards the GPT equivalent or the GLM one (I heard GLM gives you 20x the token usage for the same price of anthropics 5x)
I find it difficult to compare LLMs in general. It will do better in some cases, worse in others. As far as I see it, most of it is marketing. You can try it out extremely cheaply though through their API and OpenCode. I personally would recommend it.
this makes sense. I will give it a try via opencode and see how does it do.
GLM is Sonnet 5 level, maybe slightly below. Probably Opus 4.5 level if you want an Opus comparison.
It's a good model that gets things done but it's just not comparable to Opus and a very very far cry from Fable. It does poorly with long horizon or very complex agentic tasks, but it's great and even cheaper than Sonnet for quick sessions.
I dont buy that, mainly because in blind tests where people dont know which model was used and choose a result, this does not seem to hold true
Coding is solved once again!
Hyped! Time to start boiling lakes!
In the few minutes I had with it I didn’t notice any impressive differences beyond it complaining loudly that I can’t talk about excel cells with “yellow backgrounds”.
Fable will not code at all.
https://x.com/AnthropicAI/status/2072163884430229756
I have to admit that when it was blocked, I canceled my max plan and asked for a refund. It felt like someone took away my previous toy. So I'm happy it's back again; I upgraded to max again. Coding aside, but Claude Design is phenomenal - for both new designs and redesigning existing UIs. So my customers will face a new wave of refreshed screens all over the place in coming days ;)
Yep claude design is nice. You can also do it from CLI or your own lovable clone. I made a fork from claudable https://github.com/holoduke/Claudable What I find stong with Fable is that it can pixel perfect copy existing sites or designs. And it can do complete conversions from plain html to nuxt, angular or react apps.
This is cool, I'll give it a try with my next pet project. So most likely next week, once I'm done with Fable ;) (seriously, I haven't started as many pet projects in the last 10 years as I did in the last 12 months).
> So my customers will face a new wave of refreshed screens all over the place in coming days
How do you determine the changes to make?
Do you A/B test?
How do you measure success?
What is your product?
How many customers do you have?
My product is AuditBadger.com - it's an AI-assisted compliance management platform (ISO27001 & SOC2) that guides you through the whole process (with everything a small business might want from such software). Having a few dozen customers allows me to still care about them personally and do onboarding for each and every one of them. During those onboardings, catch-ups, or weekly calls, I see where they struggle. This is how I determine what to work on next. There's no clear measurement of success beyond user satisfaction, though they every now and then praise me a little for UI/UX improvements. With Claude Design, I've got my design system set up (also by Claude scanning the repo); I upload a screenshot of the area I'm not happy with, prompt with some additional remarks, and after a couple of iterations, I get a proposal, which is always better than what I come up with in the first place.
Site does look great. I’ve not used Claude Design. How much do you feel like you were able to give your own personal touch to it?
I ask as I see many vibe coded products that look well but are very generic. It signals vapourware to me. But honestly your site looks a cut above. Do you have a design background?
Also is the design system you mentioned public?
Thanks, I'm seriously blushing ;)
No, the design system isn't public. But only because it's a month old and I never considered opening it. I'll give it some thought.
I don't have a design background, but I ran a software house/dev shop for almost 15 years; maybe that taught me a little. And my very first client, after seeing our very first projects, said one thing: "I don't care how ugly this is, but for god sake, please make it consistent, consistency is only think that matters long-term". Those might not be his exact words, but keeping designs consistent is imo pretty important.
Regarding the personal touch, the app itself is the result of gradual evolution. It started as an HR system, which we worked on in Prograils. It even got its first semi-professional design, which evolved over the last two years (during which I learned that bootstrapping an HRMS is a very bad idea ;)). As for the website, I have to admin it - Claude Design did it. I was testing Fable 5 previously and actually decided to give CD a try. It was the result of one prompt, which gave me five proposals. One made it to the main page, and two others went to my other pet projects (which are meant to drive traffic to AuditBadger).
[flagged]
I would say this is an important part of the product-market-fit discovery process. Sure, your comment holds true for well established products with a loyal customer base, but for an earlier stage product this rapid iteration process is quite important, especially from an onboarding and ease-of-use perspective.
Yes, this. I never redesign things for sake of redesigning. But being solo developer bootstrapping my product, I don't have a luxury of testing some concepts deeply internally or publishing well-optimized components. And when I see my customers struggle, that's clearly a signal, that's something in UX is wrong. Just today we had to explain one of our customers how to proceed with our ISMS workbook (Clauses 4-10 from ISO 27001) - initial design proved to be bad approach, I now know how to change this. And Claude Design is great translating my thoughts and suggestions into something, that's consistent and better, than I could design it in predictable time.
We really need glm to be on pair with claude. A good Chinese model against a good US model would be the ultimate competition. No bullshit scaled down models with insane restrictions. Everything open and maximum speed in model improvements. If China also would manage to master high tier chips then we truly can expect 768gb machines in 4 years from now.
In the days without Fable, I find that Opus is good enough. Now it's back, but I totally don't miss it now.
Fable 5 is functionally useless not just a regression of the original Fable. totally useless not to mention when it flags your code you may have noticed in the original release the models internal reasoning was not viewable it is intros release and mines. internal thought process said I am Claude Sonnet 3.5 when I asked it what model it was while it answered I am Fable % in the chat window. And to be fair it was about as useful as sonnet 3.5
I never strongly considered open weight models before getting throttled all the time from Claude. I bought a 1 year pro subscription but my bet is I won’t renew it.
The Claude “usage” UX is very bad. At the most basic level, there is no way to know what you’re actually paying for if you buy a “pro” plan or a “max” plan. Dario will take your $200 but he will give you a secret number of tokens in return.
I have no way of knowing what the numerator or denominator are for the usage progress bars. Plus they change them all the time. There’s no way to audit it. So if there was a malicious script siphoning usage it would be really hard to detect.
Yeah I recently downgraded my subscription. The paternalism is out of control. Secret weights, secret guardrails, secret stenography, secret dumbing-down of triggers, secret token allowances.
If you use Claude Code, there are plenty of projects/tools that let you see what your usage would have cost at API rates, then you can cross-reference that to your "usage" in Claude. I run through enough token in a few days for the subscription to pay for itself.
But I agree that advertised visibility into this would be nice, so we could compare the different providers up-front.
1) There is no way to cross reference actual usage because Anthropic won’t tell you how much they think you used.
2) It’s not a nice-to-have, it is a basic accounting question of how much of a product you are getting.
3) Even if it did work, it should be a first-party feature not a third-party add on.
You can easily figure this out though.
A 20x plan gives you between 8-10B tokens per week (99+% of which are cache_read), and has been for the past ~3 months (since I started tracking it). I have not seen any quota movement despite their claims to be raising or lowering usage limits.
I mean if you want to spend $200, and get a guaranteed amount of tokens in return you should be using the API.
I agree there should be more transparency what 20x gets you that 5x doesn't.
Though I also imagine it's a tough problem when you could be using Dispatch, Cowork, design, chat, code, etc. All of which use different contexts, models, and resources. I'd argue they actually need to either simplify their offerings, or charge more for upsells (charge for cloud-based agents, upcharge for design outcomes, etc).
If products use different amounts of tokens, that’s fine. That’s fundamental to the products. But that’s the numerator, which they aren’t telling you.
I’m also saying the denominator isn’t clear. One day it might be a million. Tomorrow 2 million. The next day 800k. Who knows.
I agree that would improve transparency but it'd also lock us and Anthropic into a situation where changing that would require new plans, prices, notifications, legal, etc. I think the fact that we don't see tokens shows just how subsidized the plan is. Which is mostly the point. They want us getting used to just using the product, not tracking tokens (for whatever behavior that builds later on).
npx codeburn
Not super impressed, but I doubt my requests are getting routed to Opus -- it just doesn't seem to be as good at mathematics as it is at code (I found this to be the case last time it was released as well).
Yeah confirmed with more usage today. It seems ~ on par with (if not slightly worse than) 5.5 on math-oriented stuff.
I believe they will keep fable available, but either reduce the usage to 25% or even 10% otherwise I don't think they would have put that much effort into flushing out a system like that.
I can use it but I'm not. I used it before the ban, it's better but the speed at which tokens are burned is not justified. And this makes me happy because it means I am ok with current capabilities that will soon be matched by open-source models. Open-source harness, open models, and soon self-hosted and private with models burned directly onto chips (like Talaas is doing).
So a model named 'Fable 5' is "back". Both excited, as the previous Fable 5 I had access to for just 3 days was fantastic, and anxious as (yes, in my 1 person anecdote) the model refered to as 'Opus 4.8' was stealth nerfed over the last 2 weeks to a degree I had not experienced since the massive nerfs back in the GPT-3.x days. Fingers crossed.
Happy to see that Fable is back, was sad to see it go after getting to use it for only 1.5 days before. Add on top sonnet 5 availability and coding is looking fine again.
I asked it to do code review of some personal projects, that will only run on my system, and it keeps giving me:
This model's safeguards flagged this message. This sometimes happens with safe, normal conversations. Switched to Opus 4.8. Send feedback with /feedback or learn more
Not sure if others are having more success.
I'm trying it right now for a side project of mine, compared to Opus it is effectively better at following instructions and somehow has better "depth" when reasoning on complex tasks. However, if it will not be part of subscriptions, I will not use it anymore.
I have yet to run into a project where I felt like I needed this and was worth justifying the latent worry about it disappearing or micromanaging quota usage. Heard about massive refactors or ports that were compressed into days but I'm wondering if it's really worth the inference cost.
funnily enough for me, Fable 5 was much better at refactoring my large codebase than Opus 4.8 was - perhaps the scope of Fable is very great
I have been testing it on my security / firewall appliance (nfSensei) and it's much improved. It's actually proceeding through a complex plan Opus 4.8 proposed.
But then again I have not tested it without a prior model planning as of yet.
Note: fable would fail immediately on any security related topic prior.
i did use Fable for a brief moment the last time around, and apart from being extremely expensive, it wasn't any order of magnitude better than Opus or even GPT-5.5 for my workloads (development, prototyping, some api testing). I probably could manage indefinitely on Sonnet class models
Does it need to be an order of magnitude better?
I run a piddly little one man SaaS company that’s doing about $30k in MRR, growing at about 10% per month.
My AI coding costs are like $500 / month on average, but I’ll gladly spend 10x that to double my growth rate. I blew through about $400 in Fable usage in six hours before it got yanked, and it was money well spent. The alternative to that six hours with Fable was probably six days with Opus, higher risk of bugs, etc. And what it built was something I’d have gladly paid an engineer $10k to build in the pre-AI era.
My bottleneck is my attention, and Fable lets me go faster through my roadmap.
That is super cool. Checkout this repo which might give you super power to code quality and cut down 80% of token cost while keeping fable running longer jobs. https://github.com/syncable-dev/memtrace-public?tab=readme-o...
Did they remove that policy that forces retention of data for fable, even if you use it on AWS or elsewhere
Nope
I have the feeling it isn't as good as it was 3 weeks ago (not including the times you get downgraded). I mean it's still great, but in my memory it was better. Anyone else?
They also announced a promotional offer in here: [0]
> For a limited time, you can use our newest model—Claude Fable 5—at no extra cost as part of your subscription plan.
> During the promotional period, you can use up to 50% of your weekly subscription limits on Claude Fable 5 at no extra cost.
So it lasts for a week until July 7, 2026 at 11:59:59 PM PT. Then they will pull the plug on that.
So it sounds like a great time to roll the dice, pull that lever, spin that roulette wheel and spend as many tokens as possible, at no extra cost at the Anthropic casino's latest upgraded slot machine called Fable 5.
Available to gamblers while capacity and availability lasts. Most importantly have fun and don't blow up your budget.
[0] https://support.claude.com/en/articles/15424964-claude-fable...
So they nerfed Mythos down to Fable and then nerfed Fable down to Opus 4.8. I assume most people here are using it for coding and other technical tasks, and so are silently getting redirected to Opus.
They really dropped the ball on this one.
A frontier lab releasing the most advanced model causes them to lose customers was not on my AI bingo card.
> There's an issue with the selected model (claude-fable-5[1m]). It may not exist or you may not have access to it. Run /model to pick a different model.
Just vibes, but what if they're preparing for the government to buy stock? That way the narrative makes more sense of to dangerous to release.
I cant use it. Even basic coding redirects me to Opus. Even generating a hello world is flagged.
I'm not getting the Opus 4.8 switch for coding, supposedly given how fast I reached the usage limit, which is kind of nice.
Sucks that they are going to put it behind API on the 7th
Doesn't make any sense for codex users to switch
5.6 should be coming out any day now
Glad it's back, but the price outside of the subscription is insane. API pricing is just not for regular users.
Every prompt I give Fable is timing out / not completing.
It seems like Anthropic can’t keep up with users compute needs.
fable eats usage credits at an alarming rate. i know it says 50% faster, but i honestly feel more than twice as fast
This still is flags every query I send it due to safeguards, then switches to 4.8/Opus.
Definitely NOT a marketing ploy.
Yes, just like New Coke back in the 80s. Genius 4D marketing strategy.
Their post detailing the timeline and their actions since reinforce my belief that Anthropic is among the most trustworthy AI companies.
https://www.anthropic.com/news/redeploying-fable-5
Its like saying Pinochet was one of the least murderous authoritarian dictators. The bar is low...
Where are Chinese companies on your scale? Mistral?
You can assume they're more or less co-opted by the CCCP so how much do you trust Xi? Mistral isn't even a real player.
[flagged]
If the future of SOTA is handicaped models, then what's the point?
Writing code is already okay for current open models like GLM, QWEN.
I only need SOTA models for slop reviews, clean-ups and helping me with things I am not good at.
I have been extremelly happy with the results of cheaper models recently.
Composer 2.5 is a beast, fast and cheap. The upcoming Grok will probably be even better, since cursor helped trainning it.
Anthropic will lose market share quickly if they can't do better than this fable shit show...
I wonder who in their right mind will pay for this lobotomised junk.
> After July 7, 2026, Claude Fable 5 is no longer included in your plan’s weekly usage limits. You can keep using Claude Fable 5 through usage credits, which let you pay for usage beyond what your plan includes.
This is the real story.
This is getting served with a certificate Firefox says was created by an unknown issuer. It's possible I'm getting hit by a corporate middle-box, but then I'd expect to get this on every single host name.
twitter? are u sure it's not ur corporate?
Update: It is my company's middleware box. I'm not sure why I'm not seeing this flagged on other sites.
Update2: They're blocking all (not HN!) social media now because someone got catfished on whatsapp or something.
those MITM firewalls can have different configurations for different sites
Maybe they are blocking just Claude?
Here we go!
Might need a few more Claude subscriptions.
Anthropics marketing dept were a tad overzealous with how they positioned this model, now it's biting them back.
If this was just another model without the hype doom marketing, I don't think we'd be where we are.
What’s the fable of this story?
That didn't take long.
correction: opus 4.8 is rebadged as fable, fable is not back
This is pure nonsense. I asked it to edit one (long) html file, and it used up 5 hour session limit and it didn't even finish. For comparison, same task with Gemini Pro 3.1 Thinking took like 5% of usage. Similar with Codex 5.3.
Strangely enough, cancelled the subscription listed the reason, and after 15 minutes session limit is back to zero.
By the gods! The next 20 minutes will be the most consequential of my life ...
Now the Chinese models can start the distilling process again so I can get a better small parameter local model.
Can we just make Opus 4.8 ish models cheaper, they do exactly what I want for 99% of problems certainly in web app development. I’m not sure what Palantir/Salesforce/agencies will do, really upset for these leaches!
For how long?
A week
Was confused, thinking I somehow missed the release of the game Fable 4, only to be disappointed when I saw this is about Claude LLM crap.
It's kind of like that time "crypto" changed it's meaning to cryptocurrencies and NFTs.
Peter Molyneux continues to re-invent himself. His newest "god game" claims to have the ability to code any game you like, on demand. Just as you'd expect from an omnipotent deity!
jiezaz krsit - I just spent $50 in like 3 minutes after running out of credits. this is just depressing. a really valuable product locked behind payment that only large corps can afford that completely locks out individuals. im so over capitalism. I hate this. I can't wait for open models to really just catch up and become the norm so that all these closed models can go do whatever it is anybody does to themselves in the dark. im so sad.
I wish we would expropriate AI labs and use it for the good of humanity instead of for profits
Capitalism delivered both these US models and the Chinese models. I get that it sucks if they’re too expensive, but the reality is they cost a lot to train and serve.
Whatever that means.
[Future voice]
[dead]
[flagged]
This self-promotion is not relevant to the discussion.
Wrong, it sure is relevant because it was ported with Fable 5. Did you even read my post?
Also just a suggestion - try not to hate on others so much, it'll make your life more pleasant!
crashed when I clicked play campaign
Try picking a map from the map picker up on the top left drop down section, then click g-23 and you'll be matched against bots.