Anthropic Shuts Down Fable 5 and Mythos 5: Frontier Open Weights and Technological Sovereignty

Friday evening, Washington time, Anthropic disabled its two most capable models, Claude Fable 5 and Claude Mythos 5, for every customer in the world, and the debate over open weights flared back up overnight. Three days after launch. The cause is a US government export control directive, barring access by any foreign national, inside or outside the United States, including the company’s own foreign-national employees. Selective compliance was impossible, the switch was flipped for everyone, everywhere, and many read the block as the push open weights had been waiting for. The two models had been available for only a few days, and it’s the first time a directive of this kind has hit a US lab’s flagship models this way, in a single evening.

Ethan Mollick, on LinkedIn, wrote a prediction that runs in the opposite direction.

The block won’t bring more open-weight models. If anything, Mollick writes, we’ll see the end of frontier open weights. The reasoning is clean: if a Mythos-class model is deemed risky, not even China will have an interest in keeping it open, and you can’t build a model like that without a concentration of compute that sits inside one country, visible and regulable.

I left a comment under his post, and I want to expand it here.

On the mechanism Mollick is right, for the absolute top. On the conclusion, two corrections are needed, and both lead exactly to where I spend my days, local inference and technological sovereignty, the work I do at LocalAI.

The footprint logic holds

A model at the top of the capability curve is born today from a concentration of GPUs that fills physical data centers, inside a precise jurisdiction, with traceable power draw, supply, and contracts. The term that has been circulating in regulatory documents for a while is regulatable compute: a top-tier training run leaves physical traces, the off-the-charts electricity, the scale of the data centers, the purchase of tens of thousands of accelerators, and a state can identify and reach all of it. The same American restrictions on exporting the most advanced chips exist because that compute is visible, countable, stoppable upstream. On Friday the state did exactly that, in one evening.

An open model of seven or seventy billion parameters, by contrast, once downloaded lives a life of its own, and a copy on a laptop can’t be recalled with a directive. It’s the difference many grasped right away, whoever holds the weights locally doesn’t get them taken away by any government. Mollick anticipates the China objection, and rightly so. At the true summit, the most powerful models in absolute terms, neither bloc has any interest in letting the weights circulate freely. On this I grant him everything, the ceiling closes on both sides.

From 1.2% to nearly 30% in a year

The word “frontier,” though, in his post means the ceiling, the Mythos class. The capability that actually drives adoption sits one rung below, in the near-frontier, more than enough for almost everything companies do every day: extracting data from a contract, classifying requests, feeding a RAG system, drafting, running agents on bounded tasks. For this kind of work the distance from the ceiling has thinned to the point of irrelevance, and the most powerful model in the world doesn’t change the outcome of a field extraction or a ticket routing. And there China is accelerating, in the open.

Qwen from Alibaba, Kimi from Moonshot, GLM from Zhipu, DeepSeek, and since early June MiniMax with its M3, presented as the first frontier open-weight model that combines high-end coding, a one-million-token context window, and multimodal input. The numbers tell the shift better than any thesis. A study by OpenRouter across one hundred trillion tokens, reported together with Andreessen Horowitz, measures the share of Chinese open-source models in global LLM usage rising from about 1.2% at the end of 2024 to nearly 30% a year later. The US-China Commission’s March paper reports an estimate from an a16z partner that around 80% of American startups build on Chinese base models, and notes that between November and December 2025 seven of the ten most downloaded models on Hugging Face came from Chinese labs. Western proprietary models still lead, around 70% of total usage, but the slope of the open curve all points one way. Adoption then feeds itself, the more developers download a model the more tools, integrations, and materials grow around it, and the more that model becomes the obvious choice for the next project. It’s a flywheel that works in favor of whoever publishes the weights, and right now the one publishing them most aggressively is China.

The end of Western frontier open weights

What risks closing, then, is not the open layer, it’s its Western flank. If the United States regulates its own high rung shut and Europe keeps slipping out of the top ranks, with Mistral dropping off the leading labs, the open near-frontier layer doesn’t disappear, it changes hands. It passes to Chinese labs, which keep it open precisely because openness is a competitive lever against closed American APIs, a way into stacks all over the world while the alternative locks itself down.

The block threw into relief a contrast that several observers caught immediately. A frontier open-weight model like M3 you download and run on your own hardware, and no government switches it off from a distance, while two flagship models served from a centralized endpoint vanished for everyone in the span of an evening. The uncomfortable part of Mollick’s prediction is this: a security-driven closure of the Western summit can hand the open standard to Beijing, and give a systemic competitor the default position developers and businesses build on.

For Europe the stakes are high. If the open layer that enters the stacks becomes Chinese, the digital autonomy the continent has been chasing for years ends up resting on models developed under another jurisdiction, with another supply chain and another value system upstream. Mistral remains Europe’s most serious card, and that’s exactly why its retreat from the top ranks weighs beyond the single lab. Technological sovereignty, in this scenario, stops being a conference word and becomes a choice about which ecosystem of models you want to be able to use five years from now.

One evening was enough to switch off two models

Here comes the part I see best from my own work. For the value that counts in production, frontier capability was never the thing that makes the difference. In production the difference is made by continuity of service, and an evening like Friday’s tests it more than any benchmark. Two models gone by directive, three days after launch, with Amazon asked to revoke access across all regions, and none of the customers who had built a process on top of them had any say.

It’s a trajectory, more than an isolated episode. The Department of Defense had already labeled Anthropic a supply-chain risk, and the company has opened litigation against that classification. When a vendor finds itself caught in a tension like that, regulatory volatility stops being a hypothesis on a slide and becomes an operational variable. And there’s a second, mirror-image side, if 80% of American startups run on Chinese models that exposure can one day become the object of a directive in turn: dependence on a distant vendor is a risk whatever flag the vendor flies. The lesson companies take home this week is architectural, you need a design able to withstand the moment a downtime or a block actually arrives. I’d written about it looking at the self-hosting options with Mistral, and that reasoning today holds for anyone resting a critical process on a single distant vendor.

Routing and local inference, the layer that stays yours

The architecture that withstands all of this is model-agnostic: an abstraction and routing layer that, at the exact moment something breaks upstream, moves traffic from a frontier API to an open model running in-house. It’s what we build at LocalAI, an open-source engine that works as a drop-in replacement for the OpenAI and Anthropic APIs, so the same code that called Fable 5 yesterday can call a Qwen or a DeepSeek today on your own hardware, CPU included, without the data leaving your infrastructure, with agents, RAG, and MCP support already inside. In concrete terms you change the endpoint address and the key, not the application sitting on top.

The push toward this design doesn’t come only from those who sell local inference. VentureBeat, commenting on this very block, points to an active fallback architecture as the most resilient path, with systems designed to be model-agnostic and intelligent routing layers that move traffic from a frontier model to an open-weight fallback the instant an outage or a ban hits. In the same case Anthropic, to avoid leaving everything stalled, rerouted the surviving requests to Opus 4.8, a less capable but still-running model. It did so because when the summit goes dark you still need somewhere to fall back to, and that somewhere, if it’s truly yours, no one takes away.

Drop-in compatibility means using the same SDKs and the same call structure, and the routing layer decides request by request where to send the work, based on how sensitive the data is, on cost, and on how much you actually need the largest model’s power. An internal draft stays home on a local model, a complex synthesis can go up to the frontier cloud, and if that cloud doesn’t answer the traffic falls back to local without the user noticing. For banks, healthcare, and public administration the same engine runs on premise or in network-isolated environments, where the data simply doesn’t have permission to leave.

The real lock-in, the kind that makes migrations fail, lives beyond the model, in the embeddings, in the vector database, in the orchestration logic you’ve stitched onto a vendor. Owning the abstraction layer means being able to replace them one piece at a time, without rewriting everything. And on cost the nature of the spend changes, local inference has a predictable cost tied to hardware, more than a metered bill that grows with usage and that a vendor can adjust whenever it wants. It’s the same family of tools, from LocalAI to LocalAGI to LocalRecall, whose broader meaning I told around digital skin and autonomous agents.

Agents make the problem worse

There’s a level where all of this weighs double, and it’s the level of agents. An agent that depends on a single frontier API to plan its steps and call its tools stops dead the moment that API is cut, and it doesn’t stop one function, it stops the whole process you’d handed it. The more autonomous and embedded in the workflows the agent is, the higher the cost of a sudden interruption, because you’ve moved onto the model not a single answer but a chain of decisions.

A routing layer with local fallback is what lets an agent degrade gracefully, switching to a model running in-house and keeping working, maybe a bit slower, instead of shutting off halfway. It’s one of the reasons LocalAGI sits on top of LocalAI, agent orchestration is worth something only as long as there’s an inference underneath you can’t afford to lose from one moment to the next.

Defensibility moves from the model to control

For months I’ve been insisting on a thesis that this week finds an unwelcome confirmation. When frontier capability becomes a commodity you rent, until they switch it off on you, “having the best model” stops being a defensive moat. The asset that stays defensible is control, over inference, over where the data lives, over the routing layer that holds the rest up. Capability you rent in an afternoon, control you build, and that’s why it’s worth more.

For Italy and Europe this isn’t theoretical. The AI Act’s obligations for high-risk systems land in full on 2 August 2026, with risk management, data governance, logging, and human oversight to demonstrate, GDPR makes the location of data a legal question before it’s a technical one, and the push for sovereign cloud is already redrawing which vendors can serve public projects. A stack you own answers all three at once, compliance, data residency, and continuity, and it does so without having to hope the upstream vendor doesn’t change its mind.

In practice you start by mapping the AI dependencies you have, process by process, to know what stops if a vendor disappears. From there you introduce an abstraction and routing layer between your applications and the models, you keep a local fallback ready for critical workloads and sensitive data, and you start treating inference the way you treat a company’s energy, with a main supply and a reserve that doesn’t depend on it. None of these steps requires giving up frontier models when you genuinely need them, it only asks you not to be their prisoner.

What stays with me, now that this week is over, is not the geopolitics of compute. It’s how much of our operational intelligence already runs on a switch held by someone else. We’ve grown attached to capabilities we don’t own, that can change, expire, or be turned off from far away, and in *Pelle Digitale*, my book on digital mediation, I tried to say that the technology that extends us is also the technology that exposes us, every time we give up governing it. Owning the layer that keeps you running is becoming a daily choice, one to remake every morning instead of taking it for granted. Without a doubt, that’s the question I carry into every meeting these weeks: how much of your operational intelligence are you willing to leave on a switch you don’t hold in your hand?

Source: Ethan Mollick, post on LinkedIn, 12 June 2026.

On the facts of the block: Anthropic’s statement, CNBC, Tom’s Hardware, VentureBeat.

On the market data: OpenRouter and Andreessen Horowitz, the US-China Commission (USCC) paper.

Originally published in Italian on fabiolalli.com.