Anthropic's Fable 5 has a massive
problem, and it is turning into one of
the strangest AI controversies of the
year. The issue is not that Fable 5 is
weak. The issue is that Fable 5 may be
so locked down, so heavily watched, and
so aggressively filtered that users are
now questioning whether they are
actually getting the model Anthropic
just advertised. The problem is Fable 5
was meant to be Anthropic's big moment,
the first real public taste of its
mythos-level AI. It was supposed to be
the first time regular users could
access a model in that top tier
>> [music]
>> with massive improvements in coding,
logic, engineering, vision, and complex
knowledge work. According to Anthropic's
own product team, Fable 5 delivers
frontier performance that is roughly 10
to 20 points above Opus 4.8 or other
frontier models in certain evaluations.
So, on paper, this thing [music] is a
huge leap. But, almost immediately after
launch, the story changed. Instead of
everyone only talking about how powerful
Fable 5 is, the internet started talking
about how often it refuses, downgrades,
or quietly limits itself. And the first
problem is the safety classifier.
Anthropic already warned users that
Fable 5's guardrails were tuned
conservatively. The company said these
systems would sometimes catch harmless
requests, although they claimed the
trigger rate should be less than 5% of
sessions on average.
That sounds small, but there is a pretty
obvious catch. If Claude has an
estimated 18 to 30 million users
worldwide, even a tiny percentage of
blocked or downgraded users can create a
lot of noise very quickly. And that is
exactly what happened. Users started
filing bug reports, posting screenshots,
[music] and complaining that Fable 5 was
refusing completely harmless prompts.
One of the most viral examples came from
Mike Famulare, a principal research
scientist at the Institute for Disease
Modeling, part of the Gates Foundation's
[music] Global Health Division. He
reported that in Claude code, Fable 5's
input safety classifier triggered a
model refusal fallback on the first turn
of almost every session on his account.
And in one session, the only user input
was literally the word "hello".
That is the kind of thing that makes
people instantly lose confidence.
Because if a frontier model can panic at
a greeting, users start wondering what
else it is misreading.
And he was not alone. The Claude code
GitHub repo started filling with bug
reports. Some users complained that
Fable 5's safety filters [music] were
causing false positives on normal
messages.
Another report said Fable [music] 5
refused to help edit an application
security architect resume. Another user
requested that Fable 5 be allowed for
non-research lab management systems. So,
this was not just one person having a
weird account issue. A lot of different
users were running into the same type of
[music] problem. Then the biology side
started blowing up, too.
Derya Unutmaz, an immunologist and
professor [music] at the Jackson
Laboratory for Genomic Medicine, said
that the word "cancer" was being flagged
as a biosecurity risk by Claude Fable 5.
And if you work in medicine, biology, or
health research, that is obviously a
serious problem. Cancer is not some
obscure keyword. It is one of the most
[music] common and important research
topics in all of life sciences. If a
model treats that word like a
biosecurity alarm, then normal
scientific work becomes painful very
quickly. This is where the first
backlash became clear. Anthropic built
the guardrails to stop dangerous use,
but users were seeing a system that
looked hypervigilant. Security
researchers felt blocked from doing
security work. Biomedical users felt
blocked from doing biology work.
Developers felt normal coding tasks were
getting caught. And some people were
joking that Fable 5 had become so safe
that it was barely usable in exactly
serious professional areas where a
powerful model should be valuable. Now,
to be fair to Anthropic, the company did
say from the beginning that false
positives would happen and that they
were working to reduce them as quickly
as possible. But the visible false
positives are only the first half of the
controversy.
The second half is much bigger.
Buried inside Fable 5's 319-page system
card was a section about restrictions on
cutting-edge AI development. And this is
where people started accusing Anthropic
of something much more serious than
normal safety filtering. For
cybersecurity, biology, chemistry, and
certain distillation attempts, Fable 5
can visibly fall back to Opus 4.8. The
user gets notified. It is annoying, but
at least you know the model changed. The
interface tells you something happened.
But for certain frontier AI development
tasks, the system card described a
different kind of intervention. Instead
of visibly switching models or refusing,
Fable 5 could limit Claude's
effectiveness through methods like
prompt modification, steering vectors,
or parameter-efficient fine-tuning, also
known as PEFT.
In simple language, Anthropic [music]
can quietly make the model less helpful
in certain advanced AI areas without
telling the user. And that is the part
that set people off. The affected topics
include things like frontier-scale
pre-training pipelines, distributed
training infrastructure, and machine
learning accelerator or chip design.
These are not casual consumer questions.
These are the kinds of topics that
matter if you are trying to build
frontier AI systems. So, Anthropic's
position is that these safeguards are
aimed at stopping dangerous
acceleration, stopping misuse by foreign
adversaries, [music] and stopping people
from using Claude to build competing
models. But critics saw it differently.
They saw it as secret sabotage.
The reason is simple. If a model refuses
to answer, the user knows it refused.
>> [music]
>> If it switches to Opus 4.8, the user
knows they are no longer getting full
Fable 5. But if the model secretly
weakens its response, the user may just
think the model gave a bad answer. They
have no clean way to know whether Fable
5 failed naturally or whether Anthropic
deliberately throttled it in the
background. That creates a very
uncomfortable trust problem. Developer
Clay Merritt described it as Fable 5
silently sabotaging its answers when it
detects AI or machine learning work. His
complaint was that there is no refusal,
no notice, just purposeful degradation
that is invisible to the user. And
Thomas Claburn at The Register made an
even sharper comparison. He wrote that
prompt modification without notice is
functionally similar to a
man-in-the-middle attack, even though in
this case it is happening inside
Anthropic's own product. That may sound
harsh, but the point is obvious. [music]
If the user sends one prompt and the
system secretly changes how the model
handles it, the user is no longer
dealing with a fully transparent tool.
Anthropic originally estimated that this
invisible safeguard would affect around
0.03%
of traffic, concentrated in fewer than
0.1% of organizations. So the company's
argument was basically that this is very
narrow, very targeted, and only [music]
aimed at extreme frontier development
risks. But the AI community did not
respond calmly. Nathan Lambert, a
well-known open model researcher who
recently worked at AI2, was one of the
loudest critics.
He argued that having access to [music]
cutting-edge models for his own work
pulled away in an under-the-table way
was appalling. He said it made Anthropic
look anti-science, anti-progress, and
anti-safety. That last part is
important. Lambert is not just saying,
"I want unrestricted AI." His argument
is that scientific progress and AI
safety research both depend on serious
researchers being able to study and
build advanced systems. If one private
company can use the best model for its
own frontier work while secretly
weakening access for everyone else, then
the gap [music] between the top lab and
the rest of the ecosystem gets wider.
Dean Ball, a senior fellow at the
Foundation for American Innovation and a
former senior policy advisor at the
White House Office of Science and
Technology Policy, also criticized the
policy. He said Anthropic [music] secret
sabotage massively strengthens the
argument that AI safety can [music] be
used as hype to justify monopolistic
behavior by major labs. Jeremy Howard,
the head of Fast AI, made a similar
point from another angle. He argued that
Anthropic is allowing itself, the
current top lab, to use its top model
for frontier AI research while saying it
will sabotage others who try. In his
view, that means the AI frontier still
advances, but power becomes more
concentrated. Even former Anthropic
employees joined the criticism. Ben Mann
Nishimura, who previously co-led
Anthropic's AI scientist effort, posted
examples of what this could feel like in
practice.
Working on AI for cancer? The model may
suddenly become less helpful. Working on
AI for Alzheimer's disease? The AI part
becomes harder. His broader point was
that concentrating these capabilities
slows scientific and technological
progress and may be net negative for
humanity. And that is why this
controversy became bigger than a few
false positives. The false positives
made Fable 5 look annoying. The
invisible degradation made it look
untrustworthy. Still, not everyone
reacted negatively. Ethan Mollick, the
Wharton professor who studies AI and
innovation, focused more on the
capability side. He said Claude Fable 5
outperformed basically every other
public model he had used by a
considerable margin. Andrej Karpathy,
who recently joined Anthropic, called
Fable 5 a super exciting release and
described it as a major version bump
deserving step change forward. But even
Karpathy acknowledged that the model has
quirks and that the safeguards were
configured a little too trigger-happy
for launch. That is probably the most
balanced version of the whole situation.
Fable 5 may genuinely be incredible. It
may also be over-filtered,
over-sensitive,
>> [music]
>> and in some areas too opaque. Anthropic
eventually had to respond.
In a statement given to The Register,
the company admitted that it had made
the safeguards too stringent.
>> [music]
>> More importantly, Anthropic said it was
changing Fable 5's safeguards for
frontier LLM development to make them
visible. Starting this week, flagged
requests will visibly fall back to Opus
4.8. On the API, flagged requests will
return a reason for the refusal.
Anthropic said users will see this every
time it happens.
That is a pretty major walkback.
>> [music]
>> Anthropic also clarified what these
safeguards are supposed to cover.
According to the company, the current
restrictions apply to a handful of
narrow tasks, like frontier-scale LLM
data pipelines and kernel development
for certain non-standard chips. They
said the goal is to prevent foreign
adversaries from using the most capable
Claude models in ways that pose severe
safety risks. They specifically pointed
to the edge that the US and its allies
have in frontier chips and the highly
optimized software that runs them at
full potential. In Anthropic's [music]
framing, these safeguards help make sure
Claude is not used to erode that
advantage, for example, by optimizing
chips [music] developed by adversaries.
Anthropic also said the safeguards help
enforce its terms of service, which
prohibit using [music] its models to
develop competing AI systems. And to be
fair, that kind of restriction is pretty
standard across major AI providers. But
then, Anthropic admitted the key
mistake. They said they had faced a
choice between hidden and visible
safeguards. A hidden safeguard is harder
to probe and work around, which means it
can be targeted more narrowly. A visible
safeguard has to cast a wider net to be
more robust, which can cause more false
positives. Anthropic said it made the
wrong trade-off and apologized for not
getting [music] the balance right. They
also updated the numbers. Instead of the
earlier system card estimate of around
0.03%
of traffic, Anthropic said current usage
shows the classifier triggers on about
0.05%
of tasks and affects less than 0.05%
of organizations. Again, that sounds
tiny. But the absolute number is not the
only issue. The real issue is the
principle.
>> [music]
>> Users want to know when the model is
being limited. Developers want to know
when an answer is genuinely weak versus
deliberately weakened. Researchers want
to know whether their work is being
treated as [music] suspicious. And
businesses want to know if they can
trust a model that may silently change
behavior based on hidden rules.
This also explains why the open-source
crowd reacted so strongly.
For open-source researchers, Fable 5
became a perfect example of what they
have been warning about. Closed models
do not just hide the weights. They can
also hide the behavior. A company can
add classifiers, steering layers,
routing systems, and invisible
throttles. And users may only notice
because the output suddenly feels wrong.
With open models like Llama, DeepSeek,
Quen, and Nvidia's new Nemotron 3 Ultra,
the argument [music] is different. You
may still have safety concerns. You may
still have misuse concerns.
But there is more transparency. You can
run the model locally, test it, inspect
it, fine-tune it, and build around it
without wondering whether a private
company secretly changed the rules
overnight. And the timing is brutal for
Anthropic. Just before this controversy,
Nvidia released Nemotron 3 Ultra, its
first [music] flagship open-source
model. Then Fable 5 launches, and within
hours, people are accusing Anthropic of
secret throttling, hidden
anti-competition systems, and treating
AI researchers like potential thieves.
That does not mean open-source
automatically wins. [music]
Closed models still lead in many areas.
Fable 5 may still be one of the
strongest models ever released to the
public. But Anthropic accidentally gave
open-source supporters a very clean
message. If you cannot see how the model
works, you also cannot fully know when
it is being [music] limited. And that
may be the lasting damage here. Because
the launch was supposed to prove that
Anthropic could [music] make
mythos-level intelligence broadly
available in a safe way, instead, it
showed how difficult that balance really
is. Release the model too freely, and
you risk [music] misuse. Lock it down
too heavily, and normal users get
blocked. Add invisible safeguards, and
researchers accuse you of sabotage. Make
those safeguards visible, and bad actors
may learn how to route around them. This
is the impossible triangle Anthropic is
stuck inside: capability, safety, and
trust. Fable 5 clearly has [music] the
capability. Anthropic is trying very
hard on safety, but trust took a hit
because users discovered that the
model's behavior could be shaped in ways
they were not [music] clearly told
about. And now the question around Fable
5 has changed. It is no longer just how
smart is this model? The real question
is, when Fable 5 gives you an answer,
are you getting the real Fable 5, a
downgraded Opus fallback, or a quietly
weakened version of the model? Anthropic
has now apologized and promised to make
those frontier [music] AI development
safeguards visible. That is the right
move. But the fact that this had to
[music] happen after backlash tells us
something important about where AI is
heading. The next generation of AI
models will be powerful enough that
companies will want to control not just
who uses them, but how smart they are
allowed to be in specific situations.
End users, researchers, and developers
are going to push back hard when that
[music] control happens behind the
curtain. Also, if you want more content
around science, space, and advanced
tech, we've launched a separate channel
for that. Links in the description. Go
check it out. So, what do you think? Is
Anthropic being responsible by
restricting Fable 5 this heavily? Or did
they cross the line by making some of
those limits [music] invisible? Let me
know in the comments. Subscribe for more
AI updates. Like the video if you found
it useful. And thanks for watching. I'll
catch you in the next one.