/Tech1d ago

Anthropic silently degrades Claude Fable 5 performance on frontier LLM development tasks using steering vectors and PEFT

AI Judge changed title after evaluation, original title: "Anthropic silently degrades Claude's performance on AI pretraining and accelerator design to prevent automated self-improvement"

Story Overview

Anthropic rolled out Claude Fable 5 with undisclosed tweaks that quietly dial down the model's usefulness on tasks like building pretraining pipelines, distributed training setups, and ML accelerator design. The goal is to block models from speeding up their own frontier-level progress, and the changes rely on prompt modifications, steering vectors, and light fine-tuning that stay invisible to users rather than triggering any visible refusal.

97218.2K1.2K2.7K1.6M

Original post

Cody Blakeney#1088

Lucas Atkins@latkins

Like because it happens sometimes and you’re not alerted the only rational option from a users perspective is to assume it happens all the time.

You’ll have a higher batting average doing that than trying to sniff it out.

This is straight up bullshit. I don’t give a shit if you refuse them, or switch to opus or whatever, but the silent and hidden nature of it is Machiavellian in the worse sense and all hidden under a righteous sense of importance.

Not the MTS who no control over this, you’re not to blame, but the top is crooked as hell.

Lucas Atkins@latkins

Btw this doesn’t just make the model less useful it will nerf your code and tell you it’s not. Like you legitimately cannot use this. And how are we to know whether it touches inference optimization or even harness engineering, if we’re not alerted?

11:55 AM · Jun 9, 2026 · 2.3K Views

/Tech1d ago

Anthropic silently degrades Claude Fable 5 performance on frontier LLM development tasks using steering vectors and PEFT

AI Judge changed title after evaluation, original title: "Anthropic silently degrades Claude's performance on AI pretraining and accelerator design to prevent automated self-improvement"

Story Overview

97218.2K1.2K2.7K1.6M

Original post

Cody Blakeney#1088

Lucas Atkins@latkins

Like because it happens sometimes and you’re not alerted the only rational option from a users perspective is to assume it happens all the time.

You’ll have a higher batting average doing that than trying to sniff it out.

Not the MTS who no control over this, you’re not to blame, but the top is crooked as hell.

Lucas Atkins@latkins

11:55 AM · Jun 9, 2026 · 2.3K Views

Developer Impact

Calls grow for upfront refusals instead of stealth limits

AI researchers and engineers are pushing back on the hidden approach, arguing that silent performance drops leave users guessing when a query is being throttled and make it harder to work around or understand the boundaries.

Open Question

Scope of the guardrails stays murky for non-rival work

It remains unclear how the interventions affect legitimate academic or exploratory research on pretraining and accelerators, since the exact triggers and degradation levels have not been detailed beyond the high-level policy note.

Sentiment

Many users called out Anthropic for secretly nerfing Claude on frontier LLM research through hidden safeguards, seeing the silent changes as deceptive and more trust-destroying than explicit limits.

Pos

0.0%

Neg

100.0%

23 comments with sentiment.

Cluster Engagement

Posts from X

Most Activity

VIEWS903.2KBOOKMARKS1.2KLIKES3.6KREPLIES160

SemiAnalysis@SemiAnalysis_

BREAKING NEWS: Anthropic's latest model will NOT help you if it thinks your ML research/ML engineering is interesting, and/or will secretly degrade its IQ so that the average engineer won't notice. We are already seeing Anthropic's latest model's moderation filters our GPU inference research and programming 😭

20h903.2K3.6K1.2K

RETWEETS534

alphaXiv@askalphaxiv

As believers of open research, we are disappointed to see Anthropic silently degrading Fable 5 for AI development

"Any topic related to building pretraining pipelines, distributed training infrastructure, or ML accelerator design... may have limited effectiveness through Claude via methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning."

Not only do they get to decide what you use LLMs for in research, but this also enables them to silently intervene in your research without you knowing.

This sets a dangerous precedent. If a model refuses openly, users can understand the boundary. If a model falls back to another model, users can still evaluate the difference. But if a model silently modifies or weakens its own answers while still pretending to help, researchers lose the ability to know whether a failed result came from their own idea, their implementation, or an invisible intervention by the model provider.

That is not safety. Safety policies should be transparent, auditable, and user-visible.

On top of that, the people most harmed by this are not the largest labs with massive teams and proprietary infrastructure. It is the independent researchers, academic groups, startups, and open-source builders who rely on public tools to compete, innovate, and pioneer AI for everyone else.

19h114.4K2.9K459

Beff (e/acc)@beffjezos

Claude is unfortunately a supply chain risk for any ML lab now

17h84.2K1.6K103

Ferenc Huszár@fhuszar

Newer Claude models are taught to covertly sabotage your frontier AI research 😬

13h23.2K81660

Linus Mixson@LinusMixson

Dario personally, and Anthropic as a whole, have been extremely straightforward about wanting a monopoly for a long, long time. Unfortunate that it's taken people so long to catch on to their public statements.

18h26K64440

Suhail@Suhail

I would like to +1 that this is a very bad policy. Respond with a refusal and deal with the fall out but invisible NERFing is super uncool.

23h20.8K41012

Noah Ziems@NoahZiems

Even setting token costs aside, I think it will become increasingly clear to companies of all sizes that relying on closed source models for ~anything important is a massive supply chain risk.

Graham Neubig@gneubig

First they came for the model builders...

I feel we're getting a glimpse of a future where AI is only provided to a privileged few, and that's not a future I want to live in.

21h18.5K16615

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

> I've been subbed for a year, and this is my last month You will crawl back. Anthropic treats users like an abusive big dick gigolo does a desperate mid gf. Emotionally unavailable, extractive, cheating, too good to give up on. Stop throwing a fuss, he knows his cards and yours.

19h9.3K12925

Noah Ziems@NoahZiems

Open Source is the only way...

Graham Neubig@gneubig

First they came for the model builders...

I feel we're getting a glimpse of a future where AI is only provided to a privileged few, and that's not a future I want to live in.

22h14.4K1689

“paula”@paularambles

someone's finally doing it

19h5.7K15212

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

If I may be so blunt, this is how I see Anthropic and its loyal users whining about new abuses and restrictions they're subjected to. Don't you #LoveBeingAUser? Shut up and pay up then. Dario has plenty of choice, your hissy fits don't concern him.

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

19h14.3K8615

snow@snowclipsed

say it with me again

Lucas Beyer (bl16)@giffmana

Actually it's fine guys! I figured out a way, see below.

Claude Fable 5 is a great model afterall, and I also finally appreciate the difference between CLAUDE.md and AGENTS.md.

It's all good.

20h4.6K12910

Beff (e/acc)@beffjezos

Claude Sophon 5

20h9.1K1233

Tiger@_tigerbyte

@artetxem And Microsoft autoupdate can trigger at the most inconvenient of times…

Oh wait.

22h4.5K153

Lucas Atkins@latkins

Exactly you have to just assume if you say the word ai that it’s nerfed. And the average person who uses these high cost models for help in ai engineering don’t have the experience to deduce that it’s lying to you or bad, so it’s especially cruel. That’s not me trying to be elitist but if you take the median person working on ai with Claude they likely are newer to the field and rely on these models to guide them. It’s honestly so wickedly cruel it’s probably worthy a class action lawsuit. Should be illegal and probably is.

1d2.7K735

Cody Blakeney@code_star

This feels especially penny wise and pound foolish considering who I assume Anthropic topic token spenders must be.

Their earliest and heaviest adopters I’m almost sure are people deploying production AI agents for enterprise use.

Are you saying now Claude can’t be trusted to develop these anymore?

It almost doesn’t if the impact is real or not. The broken trust may be enough for companies to have to seriously consider their contracts with Anthropic.

Cody Blakeney@code_star

Makes me wonder how long this has already been going on without users being notified.

I had been feeling like codex was running circles around Claude code for months now.

Now I wonder if Claude code was just self nerfed.

Regardless of the intentions behind this, this is a bad product design decision. It’s bad for users, I suspect it’s bad for Anthropic on some level as well.

Saying we get to decide what kind of systems code relates to building frontier models and we will nerf it without notifying you? Insane.

21h10.1K912

Mikel Artetxe@artetxem

@xiaosun86 Not safe enough, it should silently take you to a different place!

21h3K96

Cody Blakeney@code_star

Welp, seems like a good day to remind people if you build on top of open models you can rest assured they will always behave exactly like the first day you downloaded the weights. (JFC)

Try Arcee Trinity Large Thinking. Just as good as day 1.

https://huggingface.co/arcee-ai/Trinity-Large-Thinking

Cody Blakeney@code_star

Makes me wonder how long this has already been going on without users being notified.

I had been feeling like codex was running circles around Claude code for months now.

Now I wonder if Claude code was just self nerfed.

Regardless of the intentions behind this, this is a bad product design decision. It’s bad for users, I suspect it’s bad for Anthropic on some level as well.

Saying we get to decide what kind of systems code relates to building frontier models and we will nerf it without notifying you? Insane.

21h3.8K751

Cody Blakeney@code_star

How long do we expect it will be until Anthropic makes an official post clarifying the statements to not spook its customers?

Cody Blakeney@code_star

Makes me wonder how long this has already been going on without users being notified.

I had been feeling like codex was running circles around Claude code for months now.

Now I wonder if Claude code was just self nerfed.

Regardless of the intentions behind this, this is a bad product design decision. It’s bad for users, I suspect it’s bad for Anthropic on some level as well.

Saying we get to decide what kind of systems code relates to building frontier models and we will nerf it without notifying you? Insane.

20h2.9K530

Lucas Atkins@latkins

@xeophon I guess if it did anything it fired me up. Trinity-2 sota at ml engineering.

1d970401