The argument in one line.
The Claude model generating all the headlines is locked to a government invite list — what paid subscribers actually received is a capable but deliberately constrained sibling that silently downgrades you the moment you touch the three fields it was supposedly built to master.
Read if. Skip if.
READ IF YOU ARE…
- You are a developer or AI builder deciding whether to pay for a higher Claude plan expecting Mythos 5 access.
- You work in cybersecurity, biology, or chemistry and assumed the new release would unlock better answers in your domain.
- You are on a Claude Pro or Max plan and want to know exactly what Fable 5 inclusion means before June 22.
- You saw benchmark charts shared on Twitter or YouTube and want to understand what the numbers actually measure.
SKIP IF…
- You have already read Anthropic's full announcement and understand the Fable/Mythos distinction and Project Glasswing.
- You need a deep technical benchmark comparison rather than a plain-language policy and access breakdown.
The full version, fast.
Anthropic released two models simultaneously: Mythos 5, which is invite-only through Project Glasswing, and Fable 5, which is the same underlying model with a safety classifier layered on top. For most everyday tasks Fable performs at full power, but in cybersecurity, biology, and chemistry the queries are silently rerouted to the older Opus 4.8 model. On top of the capability restriction, Fable 5 costs twice as much as Opus 4.8, imposes a 30-day data retention window with no zero-retention option, and is only included in paid subscription plans until June 22.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:42
01 · Hype vs Reality
Hook states the premise: Mythos 5 is real and powerful, but publicly inaccessible. Paid users get a different model.
00:42 – 01:15
02 · Mythos Origins
In April Anthropic built a model so capable at finding software vulnerabilities they held it back from public release and handed it only to cyber defenders and government.
01:15 – 02:26
03 · Mythos vs Fable
Fable 5 is the same underlying model as Mythos 5. Anthropic named them near-synonyms on purpose. The benchmark charts circulating on social media show Mythos/Fable scores combined.
02:26 – 03:30
04 · The Safety Filter
Three domains trigger the fallback to Opus 4.8: cybersecurity, biology, and chemistry. Users get a small notice but the conversation continues — just with a downgraded model answering.
03:30 – 04:05
05 · Asking Fable Directly
Host prompts Fable 5 to self-assess. It immediately discloses its own constraints, including the Opus 4.8 routing for cybersecurity, chemistry, and biology queries.
04:05 – 05:22
06 · Price, Privacy, Deadline
$10/1M in, $50/1M out — double Opus 4.8. 30-day data retention with no zero-retention option. Fable 5 exits paid subscription plans on June 22 with no committed return date.
05:22 – 06:30
07 · Who It's For
Right for developers shipping real code, long-running agents, and heavy analysis. Wrong for casual chat, brainstorming, and anyone in security or bio/chem research.
06:30 – 07:24
08 · Safety Wall Questions
Two contradictions: Anthropic publicly called for AI slowdown days before launching their fastest model. The UK AI Safety Institute already made partial progress cracking the safety system during testing.
07:24 – 08:00
09 · Final Takeaways
One-breath summary of the tradeoffs. Encourages viewers to decide for themselves based on their actual workflow.
Lines worth screenshotting.
- The Mythos 5 that generated the headlines is locked to Project Glasswing invite recipients only — the public received Fable 5, not Mythos 5.
- Fable 5 and Mythos 5 share the same underlying model; Fable adds a safety classifier that silently routes flagged queries to Opus 4.8.
- The three filtered domains — cybersecurity, biology, and chemistry — are exactly where the capability gap between Fable and older models is greatest.
- Fable 5 costs $10 per million input tokens and $50 per million output tokens, twice the price of Opus 4.8.
- Anthropic says over 95% of conversations never touch the safety filter, but the users most likely to hit it are the researchers the model was built to serve.
- The 30-day data retention window on Mythos-class models eliminates zero-retention options, a hard blocker for healthcare, legal, and financial workflows.
- Fable 5 is included in paid Claude subscriptions only until June 22, after which it shifts to usage-credit billing with no committed timeline for return as a standard feature.
- The UK government's AI Safety Institute made progress toward cracking Fable's safety system during a short testing window — Anthropic confirmed this in their own announcement.
- Stripe reportedly compressed a two-month code migration into a single day using Fable 5 — the developer and agentic use case is documented with named companies.
- Asking Fable 5 to honestly evaluate itself produces an accurate self-disclosure of its constraints, including the Opus 4.8 fallback behavior.
- Fable and Mythos are not separate products — they are named to be near-synonyms, and Anthropic stated flatly that Fable is the same brain with safety bolted on.
- The irony the hype cycle skips: the people who would extract the most value from Mythos-class capability are the exact ones locked out of it by the safety classifier.
What the Fable/Mythos distinction actually means for you.
WHAT TO LEARN
The benchmark scores are real, but the model those benchmarks measured is not the model you can open — and the gap between the two matters most in exactly the fields where the new capability would count.
01Hype vs Reality
- The strongest signal that an AI launch is being over-simplified is when the headline capability and the accessible capability are two different things with the same name.
02Mythos Origins
- A model withheld from public release due to dual-use risk is not the same as a model that failed testing — the capability ceiling is higher, not lower, which is why it was held back.
03Mythos vs Fable
- When a provider names two products as near-synonyms and markets them jointly, the practical differences between them are worth reading in the fine print before drawing conclusions from the benchmark table.
04The Safety Filter
- A silent mid-conversation model switch changes the reliability contract: you can no longer assume the answer came from the same capability level as the rest of the session.
05Asking Fable Directly
- Prompting a model to disclose its own constraints is a useful audit step. A model that accurately self-reports its fallback behavior gives you a fast way to verify what you are actually working with.
06Price, Privacy, Deadline
- Double pricing, a 30-day retention floor, and a two-week subscription window are three independent reasons to evaluate Fable 5 against your actual use case rather than the headline benchmark.
07Who It's For
- For long-running code tasks, complex agents, and deep analysis, the capability improvement is real and most queries will never hit the filter.
- For casual use, brainstorming, or writing, the older generation of Claude models delivers equivalent results at half the cost with no restrictions.
08Safety Wall Questions
- A safety system that an adversarially motivated government tester has already made partial progress against in a short window is better described as a defense-in-depth layer than an absolute barrier.
09Final Takeaways
- Knowing which version of a model you are actually running, and under what conditions it silently substitutes a different model, is basic due diligence before building a workflow that depends on consistent capability.
Terms worth knowing.
- Claude Mythos 5
- The full-capability Mythos-class model from Anthropic, released but restricted to invite-only access through Project Glasswing. Not available to general Claude subscribers.
- Claude Fable 5
- The publicly accessible version of the Mythos-class model. Same underlying architecture as Mythos 5 but with safety classifiers that redirect certain queries to the older Opus 4.8 model.
- Project Glasswing
- Anthropic's restricted access program for Mythos 5, initially limited to cyber defenders, government agencies, and selected critical infrastructure partners.
- Safety classifier
- A filter layer bolted on top of Fable 5 that monitors incoming queries. When a message touches cybersecurity, biology, or chemistry, it is rerouted to Opus 4.8 rather than answered by Fable 5.
- Opus 4.8
- The previous flagship Claude model. Serves as the fallback when Fable 5's safety classifier is triggered. Costs half as much as Fable 5 per token.
- Zero-retention option
- An API setting that prevents a model provider from storing conversation data at all. Mythos-class models enforce a 30-day minimum retention, making zero-retention unavailable for regulated industries.
Things they pointed at.
Lines you could clip.
00:00
“Claude Mythos 5 is being called the most powerful AI ever made. And that is true. You just can't use it.”
05:58
“The people who would squeeze the most out of this power are the exact ones walled off from it. That's the irony the hype skips right over.”
07:01
“Slow down everyone. Anyway, here's our fastest model.”
07:30
“Unbreakable isn't quite the story by their own admission.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphor
00:00CloudMatos five is being called the most powerful AI ever made. And that is true, you just can't use it. What you actually get is it's leeched twin.
00:08And a lot of videos on YouTube and x are completely glossing over very important details that Entropic themselves announce. So in this video, I'm gonna go over all of those important details and by the end you might find out why this model is not for you at least at its current form. Now a quick heads up for those of you who are part of the CLOUD paid plan, in order for you to have access to Fable five right now, it is available up until June 22 and I'll explain why that is in a little bit, but you just come here on your, uh, models and it should be available for you.
00:38Again, if you're part of the Pro Max, uh, or any of the paid plans at this point. Alright.
00:42So a quick back story before we get into the details. Back in April, Entropic built a model that they announced that was so good at finding holes in software that they decided not to release it to public.
00:54Now a lot of people took that and went on different theories said AGI is here, the world is over blah blah blah blah. But they just mentioned that at this point it's very dangerous, they called it Mitos and they said that right now we're handing it over to only the cyber defenders and the government so that way the good guys can patch things before the bad guys got this kind of power in their hands.
01:15So today, Entropic announced the release of Claude Fable five and Claude Mitos five. Now the Mitos five got dropped, it has the full power, It's still locked though. It's only by invite only.
01:27And they mentioned this that right now it's available in project glasswing. That's what they're referring to. And if we click on that, this gives you all of the details on who has access to this model and what they can do with it.
01:38So again, Mito's five is fully released but it's just locked and it's by invite only. On the other hand, Claude Fable five that also got released and the trick is that this is the same exact model underneath and tropic says that flat out loud.
01:53They said it's the . Fable just has a of it. They actually even named it to hint at this thing.
02:00Fable and mythos basically means the same exact thing but the charts that a lot of people are showing you on Twitter and YouTube, if you look closely at them, they're specifically saying mythos slash fable together. So, the scores are real for both of course.
02:16I'm not going to tell you that the numbers are fakes. Of course, they're not. They released it in their official announcement.
02:21But the catch isn't the score, it's which one of them you're actually allowed to reach. Fable has these filters that are constantly watching your messages. If you touch one of these three areas which is cyber security biology and chemistry or copying AI models and your question doesn't go to Fable, it gets handed to the older model which is Opus 4.8.
02:41Now, you get a little notice but the chat keeps going and for that answer you just get dropped the generation and in their own words, they turn the filter cautiously on purpose and sometimes it'll catch harmless requests. And tropic says that over 95% of the conversation never hit that filter.
02:58Now for those, Fable performs basically the same as the locked mythos. So for most everyday stuff, you generally get the full power model that is true, I'm not hiding that. But look at who's behind the wall, biology, chemistry and cyber security.
03:12They admit that Fable falls back to the older model on most requests on those areas right now. So if you're a bio student, a chem student, cyber security or a nurse or a general researcher in these areas, you're not in luck and you're not in that lucky 95%.
03:28You're the one getting downgraded right away. I actually went ahead and asked this so I went to Fable five right here. I said, hey, so you just got released.
03:36There's a lot of hype about you. Tell me the good and the bad about you. Right?
03:39So it talks about the good that the core claim is real, the capability Fable five is state of the art and nearly tested on benchmark strong performance over everywhere, but the bad thing the bad thing right away it says, I'm deliberately constrained.
03:52Queries, uh, queries on some topics will instead receive a response from Claus Cloth Cloth Opus four point eight. Most queries related to cybersecurity, chemistry, and biology get routed there.
04:02So it mentions it literally right there if you just ask it. Now quick three things before you start using this model. First of all, the pricing.
04:09Now for developers and, uh, API or just general use case, uh, Fable five is twice as expensive or twice as much expensive as your Opus 4.8 models. And for the developers on the API, same thing.
04:22It's $10 per million words in and 50 per per million out. And again, that's double the older OPUS models. Number two is going to be the privacy matter.
04:31So it says that the Meetos class models keeps your data thirty days for safety monitoring. They say no training on it, just delete it after, but there's no zero retention options. So healthcare, law, finance, that alone might kill it for those three particular, uh, subjects or areas.
04:48Now the third one, the deadline, and again, you can screenshot this because this is very important. If you've got the paid clock, uh, plan like the pro or the max, Fable is right now included and it mentions it right here.
05:01So it actually mentions that it's only, uh, available or included until June 22. After that, it's gonna go ahead and get pulled away and you pay with usage credits. Uh, they say they'll bring it back as a normal feature when, uh, capacity allows it but no date is given at this point.
05:19So that shiny model in your app is . Now as I mentioned at the beginning of the video that this model might not be for you and here's why. Now again, this model is genuinely jaw dropping at a few things.
05:32Stripe said that they did a code migration in a day that would have taken a team two months. It built a solar system simulation that predicts eclipses from scratch. It beat the Pokemon game using only screenshots.
05:43All the models are not capable of doing that. So again, this is not hype. It's in the announcement with name companies behind it.
05:50So it is for you if you're a developer shipping real code, you run agents on long tasks or you do a heavy analysis. For you, it's a monster and most of your questions sail right past the filter. It's probably not worth it if you mostly use it for chat, right, or brainstorming.
06:08Right? The older models do that beautifully for half the price with no filter and no countdown. And it's literally not for you if you work in security or biology research or chemistry because the version built for that is the locked one and you need an invite.
06:23The people who would squeeze the most out of this power are the exact ones walled off from it. That's the irony the hype skips right over. Now, two things sit all together here.
06:32One, days before launch, Anthropic was publicly warning that AI is moving dangerously fast calling on everyone to slow down. Then, they shipped their most powerful model ever reportedly while heading towards going public. Slow down everyone.
06:47Anyway, here's our fastest model. So make that what you will hear. Now two, and this is the detail that nobody is quoting.
06:53Anthropic says outside hackers spent over a thousand hours and couldn't fully break the safety system. Very impressive. But in the same breath, they admit that the UK government's AI safety institute has already made progress towards cracking it in a short testing window.
07:09So unbreakable isn't quite the story by their own admission. Now, is fair being that open about your own weak spot is more honestly than a lot of companies would give you. But it also means that the safety wall this whole public release rests on already has a crack starting in it.
07:23Both things are the two at the same time. So here's the whole thing in one breath. The model from the headlines is locked away.
07:29The one you can touch is brilliant, runs full power for most people, but it's leased in three areas, costs are double, keeps your data a month, and walks out of your subscription on June 22. So not bad, just not for everyone.
07:42Again, maybe not for you and hopefully now you know enough to actually decide. Let me know what you guys think in the comments below. Again, you don't have to agree with me.
07:50I might be wrong here. I'm just going off based on the different research that I did and based on my own opinion.
The Hook
The bait, then the rug-pull.
Every benchmark chart being shared about Claude's new flagship conveniently omits one line: the model in the chart is not the model you can open. What actually shipped to Claude subscribers is Fable 5 — same brain, different leash.
How they asked for the click.
VERBAL ASK
00:20product
“Learn how to make money with Claude — 7-day free trial at skool.com/aiworkshop”
Mentioned in description and as a watermark lower-third. Not verbally pitched mid-video. Light touch.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
