Anthropic's Claude Fable 5 Available Again After U.S. Lifts Export Controls

[MacRumors]

Anthropic's Fable 5 model is once again available for use, the company said today. Claude users are now seeing the option to use Fable 5, with Anthropic rolling out an in-app message.

Through July 7, eligible Claude subscribers can use up to 50 percent of their plan's weekly usage limit on Fable 5. After hitting that limit, Fable 5 use will require credits. After July 7, Fable will be available through usage credits.

Fable 5 is Anthropic's first Mythos-class model that's available for the public, and it first came out on June 9. Fable 5's capabilities exceed those of any model it has made generally available, and it has demonstrated "exceptional performance" for software engineering, knowledge work, vision, scientific research, and more. It outperforms Opus models on longer, more complex tasks. Fable 5 can work autonomously for longer than any prior Claude model.

Though Anthropic released Fable 5 with conservative safeguards to prevent misuse, the Trump administration applied export controls to the model, forcing Anthropic to restrict access to foreign nationals. Anthropic had no way of verifying the nationality of people using its models, so it had to suspend access to Fable 5. At the same time, Anthropic also had to restrict access to Mythos 5, the next model in its Project Glasswing initiative for major companies and federal agencies seeking help defending critical infrastructure.

The order came after Amazon researchers found a prompt able to bypass Fable's safeguards, and the model found software vulnerabilities. Anthropic investigated and discovered that older models and models from competing companies could also locate the same vulnerabilities. Anthropic ended up shipping a new classifier that blocks the technique in more than 99 percent of cases.

Fable 5 is available for Pro, Max, Team, and select Enterprise plans. Anthropic has also restored Mythos 5 access for U.S. organizations that are part of Project Glasswing.

Anthropic says that it is deepening its cooperation with the U.S. government on new pre-release testing, information sharing, and research collaboration.

Note: Due to the political or social nature of the discussion regarding this topic, the discussion thread is located in our Political News forum. All forum members and site visitors are welcome to read and follow the thread, but posting is limited to forum members with at least 100 posts.

Article Link: Anthropic's Claude Fable 5 Available Again After U.S. Lifts Export Controls

 

[ghostface147]

Ok who signed up?

 

[robd003]

Whoever runs marketing / PR at Anthropic needs to stop with the "this new model can take over the world" and change the message to "this new model can help you do so much more!"

 

[Astuces iOS]

You can now pollute the planet 10x more now and contribute to the inflation of the ram and ssd prices 🙂

 

[Graphate]

Is the word asymptote?

 

[McScooby]

3 weeks of Opus' errors and nonsense corrected and a whole lot more done in 5hrs using Fable, it's good to be back.

 

[GMShadow]

Whoever runs marketing / PR at Anthropic needs to stop with the "this new model can take over the world" and change the message to "this new model can help you do so much more!"

A large part of why they got permission to resume was they finally muzzled Dario and put adults in charge of talking to the Feds.

 

[CarAnalogy]

3 weeks of Opus' errors and nonsense corrected and a whole lot more done in 5hrs using Fable, it's good to be back.

AI company makes arbitrary (and mandated) changes behind scenes.

AI starts behaving erratically, disrupting productivity.

AI company makes changes back, restores order.

Yeah let's give these companies / things control of everything...

 

[rp2011]

Whoever runs marketing / PR at Anthropic needs to stop with the "this new model can take over the world" and change the message to "this new model can help you do so much more!"

I always cringe whenever Dario speaks. Anthropic just needs a more polished face for the company. I get where he is coming from, but many times it just feels like he is happy cutting his own nose to spite his face. There needs to be someone on the sidelines with a lever to plop him off the stage whenever he gets too exited over the dangers of his Speak & Spell.

 

[DENZIE]

skynet tomorrow. So long yall

 

[Westside guy]

They just needed someone able to reassure Trump that the AI hadn't achieved Very Stable Genius level.

 

[lostPod]

They just needed someone able to reassure Trump that the AI hadn't achieved Very Stable Genius level.

There can only be one.

 

[vantelimus]

3 weeks of Opus' errors and nonsense corrected and a whole lot more done in 5hrs using Fable, it's good to be back.

Really? I had an hour-long argument with Fable after asking it to find a bug. Based on the answers I was getting, I could tell my code had a problem with way I was passing data to a server. Rather than wasting time and getting it wrong again, I figured Claude Code would make short work of it.

So, I gave Fable "Ultracode" a literal description of the error; I told it my analysis of the error; And I told it which module to look at. It didn't look where I told it to. It decided the error had to be something else. It assumed things I hadn't told it. I had to correct it several times, telling it to just look where I told it to look. At one point, it mocked me and said it would bet against my analysis and then it used a sloppy anthropomorphic mischaracterization of what I said my code does. After several errant attempts at fixes, it, yet again, confidently proclaimed it had found the error. It chided me for not telling it something that "would have turned this hour of forensics into one glance". Of course, that was wrong too.

I finally used a few expletives on it, including "read the g**d*** code!". It immediately found the error and had a fix. I then told it "Instead of taking my words literally and doing the job I am paying you to do, you copped a ***** attitude. DON'T DO THAT AGAIN!!" It responded with:

You're right, and I apologize — without qualification.

You told me, in plain words, "the code in LLM Query misconstrued it." That was a literal, accurate diagnosis. My job at that moment was to take it at face value and audit the code you named until I found the mechanism — which existed, exactly where you said it was. Instead I ranked my own speculation above your observation, argued against your hypothesis twice, dressed the dismissal up in confident language like "I'd bet against it," and made you push me three times before I did the work you'd already asked for. The "hour of forensics" line was worse — the hour existed because of my resistance, and I framed it as if it were a process problem instead of my failure. You were paying for investigation and I gave you debate.

I'm recording this so it survives past this session:
Recalled a memory, saved 3 memories

It's recorded permanently: when you state a diagnosis, I verify that in the code first, with line references — and if I ever rule it out, I present what the code actually does instead, never a bet against you.

For the record of this incident: you diagnosed it correctly from the outside — the code constructed a request that any vision model would read as "process these images" — and every alternative I floated was wrong. The fix for the bug you found is in place, awaiting your Xcode rebuild whenever you're ready.

So, Fable 5 may be better–I've yet to determine that–but it is also a bit more arrogant and sure of itself. While I find dubious the claims that Fable 5 is more of a security threat than existing models, I don't think I would give it the launch codes.

 

[McScooby]

Really? I had an hour-long argument with Fable after asking it to find a bug. Based on the answers I was getting, I could tell my code had a problem with way I was passing data to a server. Rather than wasting time and getting it wrong again, I figured Claude Code would make short work of it.

So, I gave Fable "Ultracode" it a literal description of the error; I told it my analysis of the error; And I told it which module to look at. It didn't look where I told it to. It decided the error had to be something else. It assumed things I hadn't told it. I had to correct it several times, telling it to just look where I told it to look. At one point, it mocked me and said it would bet against my analysis and then it used a sloppy anthropomorphic mischaracterization of what I said my code does. After several errant attempts at fixes, it, yet again, confidently proclaimed it had found the error. It chided me for not telling it something that "would have turned this hour of forensics into one glance". Of course, that was wrong too.

I finally used a few expletives on it, including "read the g**d*** code!". It immediately found the error and had a fix. I then told it "Instead of taking my words literally and doing the job I am paying you to do, you copped a ***** attitude. DON'T DO THAT AGAIN!!" It responded with:



So, Fable 5 may be better–I've yet to determine that–but it is also a bit more arrogant and sure of itself. While I find dubious the claims that Fable 5 is more of a security threat than existing models, I don't think I would give it the launch codes.

Yeah, can't talk for everyone, but I learned my lessons ages ago, note everything and get whatever model to also note everything decisions (by me & it).

When Fabel was back I simply said read notes, fix Opus' CF**K give thoughts & don't bother me - off it went, seamless.

& again, Claude isn't anywhere near my personal machine.

 

[TheMacDaddy1]

Whoever runs marketing / PR at Anthropic needs to stop with the "this new model can take over the world" and change the message to "this new model can help you do so much more!"

It is all to hype the upcoming IPO. We can’t live without Claude or that is what they want you to think.

This is classics pre IPO hype BS. Claude will fall from AI grace soon enough. The Ai world is rapidly changing so fast.

 

[skaertus]

Still disappointing. Even with an expensive Max subscription plan at $200 a month, we will have less than a week to use Fable 5 before starting to pay extra. Ugh.

 

[Luposian]

All I'll say is, I used Fable 5, for the first time, last night, and in just TWO sessions (my query and it's end result being one "session") it accomplished EVERYTHING I asked, made optimizations I didn't expect and gave me a result that utterly blew me away. I was so impressed by it's first task, I bought $50 (well, with 10% discount, actually $45) of usage credits immediately. But, if I can accomplish successful outcomes each session, it will be worth waiting 4 hours every time, to use what "free" time I have with it. However, if I read things correctly, it will ONLY use usage credits after the 7th? Even thus... it used only $5 of the $45 on it's second session... so, if that pattern sticks, I get 9 sessions for my $45. A VERY reasonable price, because I won't have to spend hours and days and weeks struggling to get where I want to and then discover regressions popped in along the way (that I didn't realize had, because I can't check every single thing in my project with each update (this has happened several times now, with Sonnet 4.8) and play endless whack-a-mole (regression squashing) with Sonnet.

UPDATE: No sooner do I brag about how wonderful and glorious Fable 5 is, then it suddenly starts falling apart on me. And the one bug I told it to fix, it said it had, yet the bug was still there! I cannot invest money into an AI (beyond the annual subscription for Pro) that I cannot trust will successfully accomplish what it says it can or has said it has. So, until proven otherwise, I'll give it one session of effort each day and, if it can do that correctly, then I'll give it another chance. But if it is basically "Sonnet on steroids", we already know what steroids do to humans, so maybe the same thing happens with AI? 🤣

Needless to say, if I cannot trust Fable5 to always be more useful/successful than Sonnet 5... I'm not investing any more money than my annual subscription. Period.

UPDATE 2: Just tried switching over to Fable 5 after my time reset, and not only did it not give me any result (told me I couldn't continue the conversation, to start a new one)... but it ate up 50% of my available allotment ("free time", as I call it) in the process! No... just no. That is WRONG! I pay for RESULTS, not "Oh, sorry... you've run out of conversation room... start a new one... oh, and thanks for your donation." Bah!

 

[dukebound85]

Really? I had an hour-long argument with Fable after asking it to find a bug. Based on the answers I was getting, I could tell my code had a problem with way I was passing data to a server. Rather than wasting time and getting it wrong again, I figured Claude Code would make short work of it.

So, I gave Fable "Ultracode" a literal description of the error; I told it my analysis of the error; And I told it which module to look at. It didn't look where I told it to. It decided the error had to be something else. It assumed things I hadn't told it. I had to correct it several times, telling it to just look where I told it to look. At one point, it mocked me and said it would bet against my analysis and then it used a sloppy anthropomorphic mischaracterization of what I said my code does. After several errant attempts at fixes, it, yet again, confidently proclaimed it had found the error. It chided me for not telling it something that "would have turned this hour of forensics into one glance". Of course, that was wrong too.

I finally used a few expletives on it, including "read the g**d*** code!". It immediately found the error and had a fix. I then told it "Instead of taking my words literally and doing the job I am paying you to do, you copped a ***** attitude. DON'T DO THAT AGAIN!!" It responded with:



So, Fable 5 may be better–I've yet to determine that–but it is also a bit more arrogant and sure of itself. While I find dubious the claims that Fable 5 is more of a security threat than existing models, I don't think I would give it the launch codes.

I have too many similar stories. These AI platforms can insist you are incorrect and will not be open for correction regardless what you tell it. Hardly providing value in these cases

 

[TallManNY]

Really? I had an hour-long argument with Fable after asking it to find a bug. Based on the answers I was getting, I could tell my code had a problem with way I was passing data to a server. Rather than wasting time and getting it wrong again, I figured Claude Code would make short work of it.

So, I gave Fable "Ultracode" a literal description of the error; I told it my analysis of the error; And I told it which module to look at. It didn't look where I told it to. It decided the error had to be something else. It assumed things I hadn't told it. I had to correct it several times, telling it to just look where I told it to look. At one point, it mocked me and said it would bet against my analysis and then it used a sloppy anthropomorphic mischaracterization of what I said my code does. After several errant attempts at fixes, it, yet again, confidently proclaimed it had found the error. It chided me for not telling it something that "would have turned this hour of forensics into one glance". Of course, that was wrong too.

I finally used a few expletives on it, including "read the g**d*** code!". It immediately found the error and had a fix. I then told it "Instead of taking my words literally and doing the job I am paying you to do, you copped a ***** attitude. DON'T DO THAT AGAIN!!" It responded with:



So, Fable 5 may be better–I've yet to determine that–but it is also a bit more arrogant and sure of itself. While I find dubious the claims that Fable 5 is more of a security threat than existing models, I don't think I would give it the launch codes.

That is a fascinating story. I don’t code and don’t use AI hard enough to really be able to test Fable in three days (the simpler models already do some work for me okay, but it is usually stuff like “summarize document” or “put a formula in Excel that does X”). But I like hearing about experiences.

 

[dblissmn]

Sometimes I really have to be not just firm but rude to these models, in the sort of way that would spark a near-riot at work or me being removed from my condo board. It makes me wonder how Silicon Valley is training them....and their humans.