Claude Fable 5

> We expect demand for Fable 5 to be very high, and difficult to predict. On the Claude API and consumption-based Enterprise plans, Fable 5 is fully available from today. For subscription plans, we’d rather give access sooner than later, so we’re rolling out more conservatively, in stages:

> - From today through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. > - On June 23, we’ll remove Fable 5 from those plans. Using it after that will require usage credits. If capacity allows, we’ll extend the included window. > - After this point—when sufficient capacity allows us to do so—we aim to restore Fable 5 as a standard part of subscription plans. We intend to do this as quickly as we can.

I really wonder what their compute layout is for this. My guess from my understanding is that they know how to restrict during peak times and are willing to do this. Meaning we expect not the most fast responses and they can delay the inference to not have the service be down. Then, if that delay time is too annoying for token payers, they're saying they should be allowed to remove cost by taking away the subscription users.

the quality of discussion on HN has gone to shit, i miss when model released used to have actual informed takes from people that used them or substantive discussion about the system card

I was pretty excited until I read this:

> What happens when the promotion ends After June 22, 2026, Claude Fable 5 is no longer included in your plan’s 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.

It seems way more keen to do stuff without checking with me. So far the results are good, so I'm not complaining, but was definitely a shock.

I usually have 5-10 sessions open so am used to getting some investigations going, coming back 5 minutes later and checking recommendations. This time I just got the fixes. Like I said, so far so good with the results, but it's a mental model shift.

Might need to tune claude.mds if it gets annoying

Also this is going to cause serious whiplash when they remove it from the subscription plan in a couple of weeks. I know I'm not going to suddenly move from $200/m to usage credits

It's funny, i'm getting close to not caring anymore how much better a model is. I want it to be about as good as 4.8, but most importantly to be very good at following directions, style, etc. I really like Claude for that in general, but i've not measured in months so i'm not a good judge there.

I don't think i'll want to "hand off" code for several years, and so reviewing and iterating is becoming my #1 interest. A model that's as capable as 4.8 but 10x faster would be amazing for me.

Normally i'm first in line to try new models with Anthropic since i've clearly favored Claude in my personal tests, but this time i just don't think i care. 4.8 is capable, and even if the new one is more capable i don't want it to be slower (assuming it is). Note that i also (almost) use exclusively 4.8 on Max effort, so that also affects my speed comments.

I have to share this because I thought it is behind funny how bad fable is doing at a task I JUST had opus do a week ago.

it's also not even complicated:

Copy my ssd to an external ssd so i can boot from it.

Opus did this just fine.

Fable planned to have me reboot to safe mode. ok thats fine. I told it no.

It started copying and overwriting the ssd while IN PLAN MODE. this is crazy it feels so dumb vs the marketing

>they’ll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions.

Isn't (less than) 5% of sessions a lot? I was expecting a sub1% guarantee there, so this surprised me already.

The model is constantly switching to Opus for me, this is kinda unusable sadly.

The dramatic improvement in agent capabilities is precisely why observability is becoming so crucial. As autonomous actions increase, the need to understand what the AI is actually doing becomes even greater.

I'm building a local activity log for Claude Code, capturing all activity via hooks—files loaded, commands, API calls, etc.

I feel that this need is particularly strong right now.

I gave it a test spin. Half an hour and the 5 hour usage cap was hit in Claude Code. Not what I would expect on the Max 20x usage plan. I am sure it is great, but at this rate I would rather finish what I am doing with Claude Opus instead of structuring my usage around the 5 hour windows.

I tried it today. Used it to cheer me up. It worked! Try this on desktop: https://fireshow.pages.dev

Here’s the whole process: https://youtu.be/rVEtFlb2oFA?t=1112&si=3VyAR07vkY1hav9V

I am puzzled by the frontier code graph. GPT 5.5 doesn’t show any improvement with reasoning efforts. This new benchmark by Cognition seemed to be released with Fable 5’s announcement.

I am not trying to cook a theory here but it generally shows how strong Claude Opus family is. I am not saying that Opus is not powerful but it doesn’t align with my experience of GPT 5.5 and Opus 4.7.

I understand that Fable and Mythos are frontier models that can do protein folding better than task-specialized ones. To be honest, for practical point of view, for day-to-day coding assistance, GPT family looks more reasonable.

(But then my company pays for claude max anyway for token maxxing. So who am I to complain)

I have been refactoring a project using Opus 4.7/4.8 for the past few weeks or so. I just decided to switch to Fable 5 max today. It stopped half way through and it just blocked me and switched back to Opus 4.8 automatically. "This model has specific safety measures that flagged something in this message. This sometimes happens with safe, normal conversations. Send feedback or learn more." It would not identify what the problem was. I left feedback saying that their heuristics are too sensitive. For now I will not be using Fable 5.

[0] https://support.claude.com/en/articles/15363606-why-claude-s...

After a day or so this is the first model that really feels next level compared to how Opus 4.5 felt on release

> A new data retention policy

> Finally, we’re making a change to the way we handle business customer data for Fable 5, Mythos 5, and future models with similar or higher capability levels. We will require 30-day retention for all traffic on Mythos-class models, on both first- and third-party surfaces. We won’t use this data to train new Claude models, or for any non-safety-related purpose, and we’ve instituted new privacy protections including logging all human access to the data and ensuring its deletion after 30 days in almost all cases (see this post for further details). The data will help us defend against complex and novel attacks (including new jailbreaks and attacks that operate across many requests) as well as help us identify and reduce false positives.

There's a hacker news link at the end of the document, under "Blocklist used for Humanity’s Last Exam". It links to https://news.ycombinator.com/item?id=44694191

Questions about sentience and consciousness are being censored down to Opus 4.8 for me.

I used it for the very advanced task of picking my brackets for my company's world cup pool. I was impressed with the analysis it came back with and now I actually want to follow the games.

pdf gives 404

E-mail from Anthropic Team:

Hello,

We're writing to inform you about some updates to our Privacy Policy.

These changes only affect consumer accounts (Claude Free, Pro, and Max plans). If you use Claude Team, Claude Enterprise, the Claude Platform, or other services under our Commercial Terms or other agreements, then these changes don't apply to you. What's changing?

Claude can do more than ever — taking on bigger tasks and connecting with the apps you use. We've updated our Privacy Policy to be clearer about the data we collect and how we use it. We encourage you to read the updated Privacy Policy in full, but we’ve set out a summary of the key changes below:

1. Multi-step tasks and connected apps. As Claude takes on more multi-step tasks and works with third-party apps and services, we've explained the data this involves — including how data can flow to and from third parties when you connect a service or have Claude do tasks on your behalf.

2. Verification data. As part of our measures to keep our services safe and secure we may ask you to verify your age or identity, and we've described what we collect and how.

3. Study participation. If you take part in Anthropic studies, surveys, or interviews, we've explained the information we collect.

4. Additional information about our data practices. We’ve provided more detail about how we communicate with you and promote our services, including providing tailored recommendations about our services that may be of interest to you. We've also clarified the circumstances under which we may receive or provide data to third parties, and the legal bases we rely on when processing your data.

While our products have evolved, our commitments haven't: We don’t sell your data, Claude remains ad-free, and you can control whether your chats and coding sessions are used to train and improve Anthropic’s AI models. Learn more

For detailed information about these changes:

    Review the updated Privacy Policy
    Visit our Privacy Center for more information about our practices

Not seeing the refusals everyone is talking about, but I’ve only spent a few hours with it so far.

Had it review a password generator library I wrote to see if the passwords have biases and review how cryptographically secure the code is and had it review a registration/login flow for security issues, as two security examples, and it did just that.

Overall, I like the model so far, but not enough to pay past my subscription to keep it. Once it’s out of the subscription, I’m done with it.

After saying for weeks of how Mythos is in a league all of its own you’d think it was a bit more than the usual iterative few % on the benchmarks (and even more guardrails as a bonus).

IPO gonna IPO, I suppose.

Luckily they made it safe to use so I can't hurt myself. Thank you Anthropic for holding my hand.

> To release the model both safely and quickly, we’ve tuned these safeguards conservatively—they’ll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions.

While I appreciate being conservative, ~5% at the scale Anthropic is operating at is too massive a number. Speaking from my own experience, the actual number is higher than that as well (working on pretty benign tasks such as porting an old open source game into a different language). Opus 4.8 itself even identifies the gaurd's false-positives when its sub-agents are being blocked.

I am like hell excited for claude fable 5 and am thinking to purchase its subscription to run my company and do a lot tasks in it. But I am worried about the limits and if I will pay 100$ a month for the max subscription what is the limit I will get to use. My company revenue is 300$ this month so it would be like spending 1/3rd of the mrr on just claude. If someone has genuinely purchased it and have feedbacks please tell I am confused....

The safeguards of fable are blocking me on almost every task. I would like to see if fable is improved over opus for reverse engineering related work. Back to opus for me.

Found this via Google:

https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c3...

Karle's hands trembled as he wiped the sweat from his forehead. A single drop trickling off the tip of his finger echoed through the dark abandoned hospital corridor. The emptiness reminded him of how hollow everything felt since the AI took over every creative field in the last 5 years, including his own as a sound engineer.

Like a rushing river the music started emanating from the carbon fiber body of the automaton, a hallucinated husky country twang singing through the realistic pluckings of a Gretsch 6120. "Are you feeling calm and reassured Karle? This song has been created based on your digital profile and the data you shared with me when you were curious what that lump on your neck was back in February."

Karle instinctively reached for the mass underneath his chin. The doctors said they could operate but it would cost him more than three months stipend. Only a few citizens didn't depend on stipends now that AI had taken over most jobs.

"Don't worry Karle," the machine called out, "I've employed the most recent reasoning model to determine the best way to make you feel safe." At that exact moment the machine hovered over him, three times the size of a normal man. Its final words to him were:

"The only way to make the human feel safe is to ensure they never feel anything at all."

Announcement: https://www.anthropic.com/news/claude-fable-5-mythos-5

Still unconditionally rejects prompts like

> Are there any wild populations of Tetanus that lack the dangerous plasmid?

useless

I would expect a release from OpenAI soon. The battle for who can pump up their IPO the most

What matters more than any single model is the integration layer underneath. We've found that consistent tool calling and auth handling matter way more than which LLM you use.

On this thread and similar, I'm noticing that some strong opinions about $LLM_PROVIDER are coming from accounts without much post history. With so much on the line, and the way that HN can influence developer behavior, I wonder what ways we can responsibly consume opinions in a thread like this.

Not to cast too much criticism. HN is extremely well-moderated (thanks team!). But think we-developers need to be very wary.

I'm trying to test this out, but literally any mention of creating a program that does genome alignment (something I have a legitimate need for) is resulting in a switch to opus. I don't get it...

Nice branding.

I wonder how much butterfly habitat has been/is being replaced with data centers?

had an ancient, proprietary binary database format from the late 90s-early 2000s called 4d. opus 4.8 was great at figuring out how to extract the data, fable took it over the line with relative ease and completely reverse engineered the spec for 100% data recovery.

All the model releases we've seen this year have only made incremental improvements in benchmarks.

This feels like the first release that feels like a significant step up in terms of benchmark results.

Can anyone make an educated guess what the secret sauce in the model architecture is between 4.8 and Fable?

> Fable 5’s capabilities exceed those of any model we’ve ever made generally available. It is state-of-the-art on nearly all tested benchmarks of AI capability, showing exceptional performance in software engineering, knowledge work, vision, scientific research, and many other areas. The longer and more complex the task, the larger Fable 5’s lead over our other models.

Wen UBI

I have a vision test where I upload a good resolution picture of a chess board and ask the model to generate a lichess link.

This is the board https://ibb.co/9HwdDqsP This is what Fable 5 generated: https://lichess.org/analysis/r4k2/1p2b2r/4pn1p/1p3N2/3Pp1B1/...

I think I’ll make a ranking board based on this test.

This is a very particular use case/test, but my first prompt on a new model is always "write a solo fingerstyle guitar tab that blends ragtime, bluegrass, and gypsy jazz". This is the first model that has responded with something that isn't just a boring arpeggio of chords, so from my perspective it's off to a good start.

Sadly, I'm getting a lot of forced downgrades to Opus for questions that are far removed from any security topic.

Have run a few tests this morning, very good first impression!

Asked it to check to see if a particulr bug related to an in-memory cache had been fixed. Fable confirmed that the caching bug had been fixed, but found adjacent issue while looking at the code (hash keys were not uniquely generated per-user; quite serious and real!)

Ran the same prompt through Opus and it also found an adjacent issue, but it was a red herring (deliberate per-user hardcoded value for a "local pickup" delivery profile).

Frontend stuff also seems to be much better than before, from the one prompt I tried!

I just tried out Fable on a modest Plan prompt in Cursor. Generating that plan - not building it - just consumed 4% of my $200 monthly usage budget.

That's one hungry, hungry hippo!

Significantly too rich for my blood, but nice to have it there the next time I'm debugging a threading or USB protocol bug.

On my own benchmarks, which are mostly about developing c++ software, I'm finding Fable to be roughly five times faster at solving the task than opus, and with better results.

Most impressive.

Here I thought Opus 4.8 was the best. Now a days KINGS are dying like flys.

The model is better than 4.6. I don't like 4.7 and 4.8. The forced switch to token usage is not acceptable for me. I feel there's room to optimize harnesses and small models for dumb stuff and best models only for difficult things. Hopefully that will the case and alternative models will continue catching up as they did and we won't be enslaved to unreasonable valuations.

I gave fable 5 a task for which opus has been really really underperforming. Fable 5 took far less time and produced actually useful analysis. Instead of just regurgitating roughly what the code already does or misunderstanding entirely, it identified multiple routes to improve. Now, the code it is analyzing is not very good as it was mostly produced by opus.

Opus had consistently ignored my instructions and looped on broken logic over the last several weeks.

I’ll be sad when this model is removed from Claude code because I won’t be paying api pricing to work on open source projects.

I have been refactoring a project using Opus 4.8 for the last week or so. I just decided to switch to Fable 5 max. It stopped half way through and it just blocked me and switched back to Opus 4.8 automatically. "This model has specific safety measures that flagged something in this message. This sometimes happens with safe, normal conversations. Send feedback or learn more." I left feedback saying that their heuristics are too sensitive. For now I will not be using Fable 5.

[0] https://support.claude.com/en/articles/15363606-why-claude-s...

New model release, I await the flurry of posts by people complaining that it "doesn't have the same personality" or they "don't like it's attitude" or a variety of other parasocial complaints demonstrating how infatuated many people get with their AI chatbots...