---
title: 'Anthropic begged the world to stop AI… then shipped this'
source: 'https://youtube.com/watch?v=1PBRhm5ZnjU'
video_id: '1PBRhm5ZnjU'
date: 2026-06-28
duration_sec: 308
---

# Anthropic begged the world to stop AI… then shipped this

> Source: [Anthropic begged the world to stop AI… then shipped this](https://youtube.com/watch?v=1PBRhm5ZnjU)

## Summary

Anthropic, despite previously calling for a coordinated pause on frontier AI development, has released Claude Fable, its most powerful model, which has been heavily lobbomized and is being hailed by some as a 'singularity moment.' The video examines whether the hype around Fable is justified or just a marketing push ahead of a potential public offering.

### Key Points

- **Anthropic's Contradiction** [0:00] — Last week Anthropic begged AI labs to pause development, but this week they released Claude Fable, the most powerful AI model ever.
- **Claude Fable's Launch** [0:18] — Claude Fable is a 'mythos-class model' that has been lobbomized for safe consumption, with guardrails that block requests in sensitive areas like cyber security and biology.
- **Singularity Claims** [0:30] — Smart people are calling Fable their 'singularity moment,' with the creator of the programming language 'Bend' saying it exceeded his expectations.
- **Performance vs. GPT 5.5** [0:32] — Fable embarrasses GPT 5.5 on the 'Trust Me Bro' coding benchmarks, suggesting superior coding ability.
- **Google Gemini as AltaVista** [0:42] — The presenter compares the current AI landscape to 2003 when Google overtook Yahoo, with Google's Gemini playing the role of AltaVista.
- **Differences from Claude Opus 4.8** [1:57] — Fable is twice as expensive ($50 per million output tokens vs $25 for Opus) and is available on paid Claude plans only until June 22, then per token. It is fundamentally the same model as Mythos 5 but with a stricter 'muzzle' (classifier models) that redirects sensitive queries to Opus 4.8.
- **Positive Developer Reviews** [2:53] — Initial reviews from software engineers are very positive, with one developer reporting massive performance improvements in their code.
- **Hands-On Test: Horse Tinder UI** [3:11] — The presenter tested Fable for UI design for his app 'Horse Tinder'. After 20 minutes, Fable demanded more payment, but the preview showed a stunning UI with realistic SVG horses and Tinder-style swipe animations.
- **Creative Touches** [4:03] — The generated UI included creative elements like 'yay or nay' and 'strictly confidential' to impress investors.
- **Render Sponsorship** [4:10] — Render is a cloud platform for developers, sponsoring the video, that allows building and deploying apps without managing infrastructure.

### Conclusion

Claude Fable appears to be a genuinely powerful AI model, particularly in coding and UI design, but its hype may be amplified by marketing tactics and a potential IPO. The presenter's test showed impressive results, but also a paywall mid-task, hinting at commercial strategies.

## Transcript

Last week, Anthropic stood on the global
stage and begged the world's AI labs to
install a coordinated brake pedal on
Frontier AI development because they're
terrified that models are getting
dangerously close to recursive
self-improvement. But this week, they
threw the pause button in the wood
chipper and slammed the gas pedal
through the floor by releasing the most
powerful AI model the world has ever
seen. Its name is Claude Fable, a
mythosclass model that's been carefully
lobbomized for safe consumption by the
average NPC. Despite the guardrails,
smart people are calling this their
singularity moment. And Fable absolutely
embarrasses GPT 5.5 on the Trust Me Bro
coding benchmarks. Right now feels like
2003 when Google was overtaking Yahoo as
the dominant search engine. And
coincidentally, Google's Gemini is like
Ulta Vista in this analogy. We're living
through another iteration of the hype
cycle. But is Fable actually legit? Or
is it just a move to pump up the numbers
before Anthropic goes public? I've been
poking it with a stick for the last few
hours and in today's video, we'll find
out if Fable is the real deal. It is
June 11th, 2026, and you're watching the
Code Report. This week is historic. On
Friday, SpaceX will go public and a new
world order will emerge. Fang is long
gone. Gay Man is dead, and we're about
to enter the age of gay manos, which is
Mexican for gay hands. But Anthropic is
the most interesting company with the
most interesting logo in this list. And
it's interesting that they named one
model mythos as in myth and another
fable which implies that it's not real.
>> I don't give two but I am telling you
right now that mother that motheruck
back there is not real.
>> It's almost like they're trolling us
with these names like that time Neil A
went to the moon or when Sam Bankman got
fried or when Anthony Weiner showed
people his we don't need to go there but
you get the point. Oh, and a quick side
note. As Sam Bankman Freed was an early
investor in Anthropic and owned an 8%
stake at one point. Had he given up on
the crypto grifting and just sat on that
investment, he'd be one of the richest
people alive right now and not in
prison. Hindsight is 2020, but how does
Claude Fable differ from Claude Opus
4.8, which was just released a few weeks
ago itself? Well, first of all, it's
twice as expensive at $50 million per
output token or no, $50 per million
output tokens versus $25 for Opus. But
another caveat is that if you have a
paid claw plan right now, you can use
Fable until June 22nd, after which point
it will be removed and can only be used
on a per token basis. That's a genius
marketing move to get people FOMO
subscribed to Claude. But is it really
that much better than Opus? Well, Fable
5 is literally the same underlying model
as Mythos 5. The only difference is the
muzzle. It has a set of classifier
models that will watch every query you
make. And if you wander into cyber
security, biology, chemistry, or of
course, model distillation, your request
gets nuked and answered by Claude Opus
4.8 instead. And that means Chinese
models like Deep Seek and Kimmy won't be
able to give us a free and open-source
Fableclass model anytime soon. Now,
among software engineers, the initial
reviews are very positive, but one of
the strongest endorsements I've seen
comes from the guy who created Bend, a
programming language for GPUs. He called
this his personal singularity moment
after Fable ripped through his code and
implemented massive performance
improvements. There are many other
stories like this, but I was skeptical,
so I tried it out for myself. I was
recently rejected by Y Combinator trying
to get funding for my app, Horse Tinder.
They said the idea was brilliant, but
the UI would never work because horses
don't have fingers. Well, let's find out
if Fable can create a better UI than a
human with 20 years of web development
experience. I put Fable on max effort
mode and told it to come up with a UI
that looks like it was created by an
elite designer. After about 20 minutes
of working, I was shocked by the result.
I got a pop-up window demanding I pay
more money for Fable to finish the job.
I felt like I was getting scammed by a
developer on Elance, but luckily we
could see some of its work in progress.
It created this absolutely gorgeous UI.
And one of the most impressive things is
that the SVG horse actually looks like a
horse and not just a bunch of random
shapes. AI typically has a hard time
generating vector graphics, but all the
icons in this UI look great, and it even
nailed the Tinder style drag animation
when swiping left or right. It also has
some nice creative touches like yay or
nay, and even put the word strictly
confidential at the bottom to make this
app seem special to investors. I'm
highly confident this will make YC 2027,
especially when combined with Render,
the sponsor of today's video. It's a
cloud platform for developers that lets
you ship production-grade apps and
agents without managing infrastructure.
That's because it runs your whole stack
in one place with your web services,
Postgress, cron jobs, and workers all
wired together by default. So you only
need to worry about one dashboard and
one mental model. You start by just
connecting a repo and pushing your code.
Then render will build and deploy
automatically. It autoscales with you up
to millions of users with a full
observability stack built in. And if
you're building an AI app, Render
Workflows gives you an execution engine
that will automatically retry when
something fails, so your agents don't
crash and burn on a long-running task.
Over 6 million developers use Render
today, and the first 2,000 people to
enter this promo code will get $50 in
free credits. This has been the Code
Report. Thanks for watching, and I will
see you in the next one.