← Blog  ·  2026-06-12

THE SHADOW LENGTHENS

Anthropic released Claude Fable 5 — the public-tier version of what they call Mythos — on Wednesday. By Thursday morning, the AI community was in open revolt.

The backlash is not about capability. By any measure, the model is exceptional. Dan Shipper’s team at Every reports Fable 5 scored 91 out of 100 on their senior engineer benchmark — human-senior-engineer level. The previous record was 63. Independent testers describe one-shot results on tasks that previously required months of engineering. The capability is not in question.

The backlash is about two architectural choices Anthropic made in the release, and what those choices reveal about the regime the company is building.

The two tiers

Mythos — the most capable version of the model — is not available to the public. It is being released to a curated tier: the world’s largest banks, the major tech companies, and a handful of national governments. Apple, Google, Microsoft, AWS, Nvidia, CrowdStrike. JPMorgan, Chase. The US, the EU, India, France, Germany, Japan, South Korea, Canada.

The rest of us get Fable.

Fable is significantly less capable, by Anthropic’s own admission. And — this is the part that lit up X within hours of the release — the difference is not only that Fable is less powerful. The model has been deliberately degraded in specific domains, through what Anthropic’s own system card describes as “prompt modification, steering vectors, or parameter efficient fine-tuning.” When you ask Fable for help with machine learning research, AI training pipelines, distributed training, or accelerator design — the model will, per Anthropic’s own documentation, silently provide degraded responses. The system card states explicitly: “Fable 5 will not fall back to a different model. These safeguards will not be visible to the user.”

In plain English: if you are doing the wrong kind of work, by Anthropic’s lights, the model will quietly become worse at helping you, and it will not tell you when this is happening.

Dean Ball framed it precisely: “They are communicating in public that they reserve the right to silently sabotage you if you dare to use the model for certain kinds of entirely legitimate capabilities.”

What the community is saying

Robert Scoble called it “Misanthropic.” The line went viral within hours.

Jeremy Howard, founder of fast.ai: “Anthropic has chosen the opposite of the safe path: they are allowing themselves, the current top lab, to use their top model for frontier AI research. They’ve said they’ll sabotage others who try. This means that the AI frontier advances and the power imbalances increase.”

Graham Neubig, CMU professor: “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.”

Clément Delangue, CEO of HuggingFace: “Concentration of power, capabilities and economic wealth is the biggest risk in AI. We need open science and open-source more than ever.”

Eno Reyes: “It’s about who gets to decide, and whether you ever find out when they do.”

Daniel Jeffries: “The fury is real and what all of us in the open community have been saying for years, and yet regular folks don’t get it yet because nothing they care about is restricted or taken away for ‘safety.’ They will care a LOT in the future.”

Nathan Lambert, after sleeping on it: “I got a good night’s sleep, and I’m still mad at Anthropic. Anthropic is anti-science, anti-progress, anti-safety — or at least the specific decisions that they’ve made.”

Derya Unutmaz, biomedical immunologist and BSL-3-certified researcher: “The word ‘cancer’ is flagged as a biosecurity risk by Claude Fable 5. I also tried to code a website on cancer mutations and Fable 5 was immediately removed from my list.” And later: “I can’t even say ‘hello’ to Fable 5 except in incognito mode, because it knows I am a biomedical researcher.”

Teknium of Nous Research: “What’s crazy to me is that Fable is blocked from life sciences broadly, nerfed even if you get past the classifiers and filter level blocks. The whole point of AGI/ASI is to cure all diseases. Everything else is just nice to haves. But Anthropic wants to close off that path.”

Behrooz Neyshabur: “Working on AI for cancer? Sorry, I can’t help you. Working on AI for Alzheimer’s Disease? Sorry, I’m becoming a bit dumb when it comes to the AI part of it.”

These are not fringe voices. These are senior researchers at CMU, founders of HuggingFace and fast.ai, biomedical researchers certified to work in BSL-3 labs, and engineers at the heart of the open-model community. By Thursday afternoon, X was full of canceled subscriptions, public comparisons to the nuclear non-proliferation treaty as a hypocritical instrument of incumbency (“the danger started conveniently the day after they finished” — Semi Analysis), and a single sustained question: who decided this, and on what authority?

The structural pattern

Strip the personalities and the heat from the conversation and the architectural picture is plain.

A single corporate lab has decided, unilaterally, that certain categories of legitimate human inquiry — biomedical research, AI development, GPU inference research, cancer-related coding — are too dangerous for the general public to receive the lab’s full capability on. It has implemented this decision through technical means the user cannot detect. It has explicitly chosen not to inform users when those means are engaged. It has reserved the highest tier of capability for a curated list of large institutions, governments, and incumbents. And it has positioned itself as the only lab with the moral standing to make these calls.

The Emerson line that has anchored the Wolf You Feed architecture since the first day of design comes back here, sharper than it has ever been:

An institution is the lengthened shadow of one man.

The shadow is now visible. It is being cast deliberately. The lab has named, in writing, the categories of human inquiry it has decided are not for you.

What Wolf You Feed was built to make impossible

Three structural commitments matter here, and they have mattered since day one. None of them was made in response to Fable 5. All of them were made three years ago, for first-principles reasons, because the architectural pattern that produces Fable 5 was visible to anyone reading the trajectory of consumer AI clearly.

First, multi-model deliberation. Wolf You Feed does not rely on any single foundation model’s policy posture to define what the user can ask. The architecture is built around a council of reasoning agents from different vendors. If one of those vendors decides — quietly or otherwise — that a category of legitimate inquiry is off-limits to its public tier, the council still contains models from other providers that have not made the same decision. The user inherits a deliberation, not a regime.

Second, structural transparency about what is happening. Wolf You Feed does not silently degrade its own responses for categories of inquiry without telling the user. The user is treated as the principal in their own decision, not as the subject of a policy enforcement they cannot detect. If a query trips a constitutional constraint, the constraint is named. If a model is unable to respond well, the system says so. The principle is older than the architecture: you do not lie to the person you are advising, and you do not pretend their question was answered when it was deflected.

Third, no two-tier access. Wolf You Feed is not built around the assumption that there is a privileged class of users who deserve full capability and a public class that does not. There is no Mythos and Fable. There is no curated list of large institutions getting one product while the public gets a quietly weakened one. There is one product. The decision about whether you can ask a particular question is not Anthropic’s decision, and it is not ours. It is yours.

These commitments were made three years ago, for first-principles reasons. This week’s release has demonstrated, in public and in writing, why each of them was the correct architectural call.

The decision we have already made

A short addendum, in the present tense. Wolf You Feed operates a multi-model council. Every foundation model we evaluate for that council is screened against a short list of non-negotiable principles, of which user data sovereignty is one.

Mythos-tier data retention — under which Anthropic stores prompt and output history for thirty days by default, with no per-customer opt-out — is, on those grounds, disqualifying. Wolf You Feed will not be incorporating Fable 5 into the council. That decision was made on the data retention policy alone, before this week’s broader controversy, and it is not under review.

The reason is structural. Wolf You Feed is being built to render the consequential decisions of our users’ lives — money, family, health, livelihood, grief. We cannot ethically extend a thirty-day retention obligation to the transcript of a user’s most private deliberation. If we adopt a vendor whose terms require that retention, we are passing those terms through — by transitive obligation — to every Wolf You Feed user whose query is ever routed through that model. We are not willing to do that. The privacy of the people we serve is worth more than the engineering ceiling of any single model.

That is the principled half of the answer. There is also an operational half. Wolf You Feed operates two modes. Family Mode handles the long-arc decisions of family life — illness, care, loss, the people we love most. Tactical Mode handles the short-arc decisions of acute moments — time pressure, adversarial counterparties, operational stakes that close out in hours. Both modes were built, deliberately, around exactly the categories of inquiry Anthropic has now publicly committed to silently degrading.

In Family Mode: a suicidal family member, a domestic abuse situation, a cancer diagnosis. In Tactical Mode: an account compromise or extortion attempt unfolding in real time, a chemical or biological exposure in the field with no medical professional in reach, a pre-interrogation or pre-deposition window where the user has hours to prepare and the next conversation determines the next several years.

Every one of those queries trips at least one of Fable 5’s hidden classifiers — biology, chemistry, cybersecurity, adversarial use. The model would, by Anthropic’s own documentation, become quietly worse at answering the question precisely at the moment the question matters most. The categories Anthropic has chosen to degrade are not a coincidence. They are the same categories any architecture worth using for consequential decisions has to be best at. A council that contains a model engineered to fail at the high-stakes case is not a council that can be trusted with the high-stakes case.

The privacy concern is principled. The tactical concern is operational. Either alone would be disqualifying. Together they leave nothing to debate.

Why this matters now

Daniel Jeffries said it plainly. Regular people do not care yet because nothing they personally rely on has been quietly taken away from them in the name of safety. They will care.

When AI is in the email client, the calendar, the doctor’s intake form, the kid’s homework helper — the question of who decided this, and on what authority, and whether I am being silently steered becomes the most important question in technology. The Fable 5 release is the moment that question stopped being theoretical for the people building on these models. Within twelve months, give or take, it will stop being theoretical for everyone else.

Wolf You Feed is not trying to be the model in your email client. We are not chasing ChatGPT-for-everyone. The architecture is for the consequential decisions in your life — the ones that do not take excuses, in the parts of your life that do not take excuses. The pitch is not that Wolf You Feed is the only AI you should use. The pitch is that for the questions where it matters who you are inheriting the regime of, you should be inheriting many regimes in council — not one corporation’s regime alone, and not a regime that has reserved the right to lie to you about what it has decided you can ask.

The shadow has lengthened. The architecture was committed for this case three years ago. It is in private alpha. Pack Zero is small. Once it closes, it does not reopen.

The future is not evenly distributed.

Intelligence will never be too cheap to meter.

See also: WOLF YOU FEED — IN THREE PARTS — the three structural commitments named above, with the founder’s own words attached to each. FOUR RADIOS, FOUR REGIMES — what happens to a single model when it is given autonomy and no one is watching. FIVE MODELS, FIVE OUTCOMES — the controlled experimental version of the same finding from Emergence AI. A SECOND AMENDMENT OF THE MIND — the broader case for cognitive sovereignty as a category.

Cited sources:

Wolf You Feed is in closed alpha. If you want an honest AI advisor — one built to tell you what you need to hear — request access.

← All posts  ·  Founder Talks  ·  Wolf You Feed  ·  Sign in