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.

Mythos somehow leaking and becoming usable by all humanity in a self-hosted manner would probably be the optimal long term outcome.

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.

I don't think Anthropic is claiming it's world ending? Just that it has offensive cybersecurity abilities which can be dangerous

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.

Most likely it'll try to rewrite the whole world's CSS and get the custom properties wrong.

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.

It's probably a huge file though, I would guess it's at least a multi terabyte file.

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.

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.

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.

Same problem, in-progress book about language and thermodynamics gets flagged. Their classifier is just a regex I guess?

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.

Won't using it eat up the whole quota immediately forcing you to pay API prices anyway?

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

Here's to OpenAI and a chinese firms bringing some much needed competition.

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.

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.

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.

Yes, I've been using Opus to write a plan and fanout sonnet subagents to implement it. Cheaper and faster

This is the goal behind Devin Fusion, pretty good results so far I think.

https://cognition.com/blog/devin-fusion

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.

this is the idea of opusplan https://code.claude.com/docs/en/model-config#opusplan-model-...

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.

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?

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.

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.

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.

That didn't take long...

  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.

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.

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.

A missed opportunity to use the HTTP 451 status code.

https://en.wikipedia.org/wiki/HTTP_451

Is refusal something that can happen mid-stream, after status and headers have been received? I haven't looked at the API

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.

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.

They didn't reset the usage either! Good luck!

It was not.

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.

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.

> 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?

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.

>So my customers will face a new wave of refreshed screens all over the place in coming days ;)

Why would you EVER think this is a positive thing? When have customers ever be happy about UI changes? UI changes are stupid. Animations are stupid. If it works, then just leave it and don't hop onto the next shiny rounded corner design. Constant design changes are annoying as hell and it sounds like you are the reason why the internet is so incredibly frustrating to use

It cannot be used for ANY coding tasks, lol.

https://x.com/AnthropicAI/status/2072163884430229756

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.

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.

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).

npx codeburn

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.

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

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.

> 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.

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.

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

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

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.

Pretty much all of OpenAI is on vacation this week.

L O L

Its like saying Pinochet was one of the least murderous authoritarian dictators. The bar is low...

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)

Haven’t played with alternative harnesses. What’s broken in your opinion and what benefit brings pi for you?

How is the harness broken?

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

twitter? are u sure it's not ur corporate?

Nope

Yes, just like New Coke back in the 80s. Genius 4D marketing strategy.

A week

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!

[Future voice]