Anthropic’s Fable 5 backlash, AI nationalization talk, and why open models matter

Anthropic’s latest flagship model, Fable 5, is one of the most capable large language models on the market. It also just became the center of a major backlash over privacy, censorship, and how much control AI labs should have over what users can do.

In a wide-ranging discussion, several tech investors and founders unpacked what Anthropic is really doing with Fable 5, why developers are furious, how this ties into calls to nationalize AI, and why all of this is quietly pushing serious companies toward open-source and even Chinese models.

What Anthropic’s Fable 5 actually changed

Fable 5 is Anthropic’s new “mythos-level” model, positioned at the very high end of capability. It reportedly tops most public benchmarks and is priced at roughly twice the token cost of Anthropic’s previous top model, Opus 4.8, with the argument that it will use fewer tokens overall because it’s more capable.

But the controversy isn’t about speed or quality. It’s about how Anthropic is controlling access and monitoring usage:

• Mandatory data retention: Anthropic is storing all prompts and outputs sent to Fable 5 for at least 30 days. That includes not just the question you type, but all the context you pass in: files, agent memories, and internal documents.

• Frontier research detection: If Anthropic’s systems detect that you’re doing “frontier AI research” (for example, using their model to design or train a competing model, or doing advanced ML/chip design work), they can silently downgrade you to a weaker model.

• Hidden downgrades and prompt rewriting: Early documentation suggested Anthropic would sometimes rewrite prompts in the background and nerf answers without telling users, while still charging for the premium model.

Developers discovered this buried in a 300+ page technical report. After a wave of backlash, Anthropic told Wired it would make these safeguards more visible and at least disclose when a downgrade happens. But the core behavior—profiling users, classifying them, and selectively restricting access—remains.

Why developers and enterprises are so angry

The reaction from the developer community has been unusually intense. The core complaints fall into three buckets.

1. Surveillance and profiling by a private company

Anthropic has publicly positioned itself as skeptical of government surveillance. Yet, with Fable 5, it is:

• Retaining all prompts, outputs, and context for 30 days

• Using that data to build a profile of each user or organization

• Classifying users into risk categories that determine what capabilities they see

For individual users, this feels like a censorship risk. For enterprises, it’s a governance and business risk: a scientist, engineer, or executive could unknowingly trip an internal safety classifier and suddenly lose access to a critical capability, with no clear recourse.

2. Silent nerfing and unequal access

Developers shared examples of seemingly benign queries triggering downgrades: asking about mitochondria, or about cancer risk and GLP-1 drugs, resulted in being kicked off Fable 5. In some cases, the system explicitly said it was switching to a weaker model because the topic was flagged as biology or cybersecurity.

This creates a new class system of “AI haves and have-nots,” where:

• Anthropic decides who is “worthy” of full capability

• Some users get nerfed answers without realizing what’s missing

• Even enterprise customers with zero-retention contracts can’t opt out for mythos-level models

Critics argue this is not just paternalistic but also anti-competitive: if Anthropic can quietly downgrade anyone doing advanced ML or chip research, it can disadvantage potential competitors while still selling them access.

3. Regulatory capture in the name of “safety”

Alongside the Fable 5 rollout, Anthropic’s CEO Dario Amodei published another long blog post arguing that transparency isn’t enough and calling for a new federal regulator—an “FAA or FDA for AI”—to approve frontier models.

Investors who have followed Anthropic for months see a pattern: warn loudly about catastrophic AI risk, propose heavy centralized regulation, and then implement aggressive internal controls that only large, well-funded labs can realistically comply with. That, they argue, is a sophisticated regulatory capture strategy designed to:

• Lock in Anthropic and a few peers as the only legal frontier model providers

• Make open-source and smaller competitors look “reckless” or illegal

• Gatekeep access to advanced capabilities for everyone else

This is why some critics now see Anthropic, not Google, as the most aggressive would-be AI gatekeeper.

How safety restrictions are pushing companies to open-source

One of the most interesting parts of the conversation came from an operator running a genomics and agricultural biotech company. For the last couple of years, his team has used large language models to accelerate highly specialized work:

• Designing RNA guides for gene-editing tools

• Predicting which gene variants might drive desirable traits in plants

• Designing genetic constructs to express specific proteins

Until recently, frontier models made this work dramatically faster and cheaper than traditional wet-lab workflows. But over the last few weeks, those same models have begun refusing to help with core genomics tasks, flagging them as potential “bioweapon” risk.

The result: his company is now actively moving away from closed US models and toward open-source models they can run locally. And they’re not alone—he says startups and large enterprises across biotech and other fields are doing the same.

The China problem: better open models abroad

Here’s the geopolitical twist: the best open-source models today, especially for some technical domains, are increasingly coming from China. US open models exist, but many are behind on capability.

So when Anthropic and other US labs voluntarily tighten their own safety policies—and when US politicians echo their doomsday rhetoric and push for stricter rules—it unintentionally:

• Pushes serious builders toward Chinese open-source models

• Accelerates China’s lead in biotech, materials science, and industrial AI

• Weakens US economic and national security over the long term

The genomics founder expects his own company to eventually build a specialized “genome language model” by fine-tuning an open base model on proprietary data. Others will do the same in their verticals. That’s great for innovation—but it also means more critical work will happen on infrastructure the US government doesn’t control.

Where to put the safety guardrails: tools vs. outputs

Everyone in the conversation agreed on one thing: AI can be used to create weapons. The disagreement is about where to intervene.

One view, associated with Anthropic’s current strategy, is to restrict access to the models themselves—especially for topics like cyber, biology, and advanced ML. That’s why Fable 5 aggressively flags and downgrades queries in those domains.

The opposing view is that this is the wrong layer to regulate. The same capabilities that can help design a bioweapon can also cure cancer, improve crops, or create powerful productivity tools. So instead of broadly nerfing the models, we should:

• Enforce existing laws on weapons, hacking, and bioweapons more aggressively

• Put guardrails at the point where digital designs become physical reality

• Use targeted screening and KYC-like mechanisms for high-risk activities

A concrete example: DNA synthesis screening

One promising model is already in place in biotech. Since 2009, the International Gene Synthesis Consortium has run a voluntary system where labs that manufacture synthetic DNA/RNA:

• Screen every ordered sequence against a database of known pathogens and dangerous constructs

• Flag and investigate suspicious orders

Major AI labs and synthesis companies recently signed a letter supporting making this screening mandatory. After 15 years of practice, they say it’s fast, automated, and doesn’t slow down legitimate research.

This is the kind of “downstream” safeguard critics of Anthropic’s approach prefer: keep models broadly useful, but tightly monitor and regulate the physical instantiation of dangerous designs.

Are AI labs inviting nationalization?

While Anthropic and others warn about AI destroying jobs and destabilizing society, politicians are starting to ask a simple question: if AI is built on humanity’s collective knowledge and is going to put millions out of work, why should a handful of private companies own all the upside?

That’s the backdrop for a recent proposal from Senator Bernie Sanders, who argued that “AI is a public resource” and floated the idea of a one-time 50% tax on the stock of major AI companies like OpenAI, Anthropic, and xAI. Those shares would go into a national sovereign wealth fund, giving the public voting rights and board seats.

Even investors who strongly oppose outright confiscation admit they understand where the politics come from. For years, AI leaders—including Dario Amodei and, in a different framing, Elon Musk—have said things like:

• “50% of entry-level knowledge jobs could disappear in 1–5 years”

• “We’ll have very high GDP but very high unemployment”

If you tell the public that you trained on all their data for free and will now use it to take their jobs, they will naturally ask for a share of the gains—or for you to be heavily regulated or even partially nationalized.

Some panelists went so far as to say AI CEOs have effectively “asked for it” by repeatedly describing their own work as dangerous and socially harmful.

A better way to share AI’s upside: sovereign wealth, not seizures

Instead of seizing equity from specific AI companies, one proposal is to reform the US Social Security system into a true sovereign wealth fund that can invest in equities, including AI and other high-growth sectors.

Today, the Social Security trust fund effectively holds a single special Treasury bond—about $4 trillion the federal government owes itself. It can’t own stocks. The idea is to:

• Convert Social Security into an account-based system where each citizen has an individual balance

• Allow the trust fund to invest a portion of contributions into diversified equity portfolios, including AI leaders

• Let all Americans become partial owners of the companies driving future growth

This is similar to how countries like Norway, Canada, and Australia run large, professionally managed public funds. It would let citizens benefit from AI-driven productivity gains without setting the precedent of confiscating specific companies’ shares.

Some participants even argued that if the US government is going to underwrite the massive infrastructure AI needs—power, data centers, critical materials—it has a legitimate case to own a meaningful equity stake in the sector, negotiated up front rather than seized after the fact.

Is AI really destroying jobs?

Another key disagreement is whether AI is actually causing mass job loss. On the ground, several operators report the opposite:

• They’re hiring more engineers and product people, not fewer

• AI is mostly boosting the revenue side—letting small teams build more products, faster—rather than dramatically cutting the cost side

• Productivity gains are creating new lines of business and more work, not less

Recent US jobs data backs this up: unemployment is low, construction and software jobs are growing, and there’s no sign yet of the 50% white-collar job wipeout some AI leaders predicted. Even Sam Altman has recently walked back earlier comments about imminent mass job loss.

That doesn’t mean there won’t be disruption, especially in specific roles. But the narrative of AI as a pure job-destroyer is increasingly out of sync with both macro data and what many founders see inside their own companies.

Closed vs. open AI: what’s at stake

Underneath all of this is a fundamental question: who will control advanced AI—centralized labs with heavy guardrails, or a diverse ecosystem of open and specialized models?

Closed labs like Anthropic, OpenAI, and others offer incredible capability and convenience, but they also:

• Control pricing and access

• Can unilaterally change safety policies and nerf capabilities

• Are increasingly lobbying for regulations that would raise the barrier to entry for competitors

Open-source models, by contrast, are messier and harder to control, but they:

• Let enterprises run models on their own infrastructure, with their own governance

• Enable domain-specific fine-tuning, like genome language models

• Prevent any single company or regulator from becoming a universal gatekeeper

Some investors are now seriously considering building massive, open compute infrastructure—multi-gigawatt data centers dedicated to open models—to ensure there’s a deep, liquid market for non-proprietary AI. But the capital requirements are staggering: a single gigawatt data center can now cost on the order of $100 billion.

That’s why the regulatory direction matters so much. If governments adopt Anthropic-style narratives and clamp down on open models, we could end up with a monopoly or duopoly of heavily surveilled, nerfed AI in the US—while the rest of the world, including China, races ahead with more open systems.

Where this leaves builders and users

For developers and companies trying to decide what to bet on, a few practical takeaways emerge:

• Assume closed labs can and will change the rules. If a capability is critical to your business, you need a plan B that doesn’t depend on a single vendor’s evolving safety policy.

• Invest in open-source literacy. Even if you mostly use APIs today, understanding how to run and fine-tune open models locally will be a strategic advantage.

• Separate safety rhetoric from actual risk. There are real dangers (especially in bio and cyber), but not every “safety” proposal is about safety; many are about market structure.

• Expect politics to catch up. As AI wealth concentrates and CEOs keep warning about job loss, proposals like sovereign wealth funds, special taxes, or forced open-sourcing will only get more popular.

If you want a deeper dive into how Anthropic’s model actually performs in practice, we spent a week testing it in real workflows in this hands-on Fable 5 review. And for a broader look at how these power struggles between labs, governments, and platforms are shaping up globally, see our breakdown of Elon Musk, OpenAI, and the global AI power struggle.

What’s clear is that the fight over AI isn’t just about model quality anymore. It’s about who gets to decide what you can build, what you can ask, and how much of the upside you’re allowed to keep.