Back to blog
17 min read

Fable became mythic: when frontier models become strategic infrastructure

Why Fable 5 vanished overnight, why OpenAI now faces a preemptive government checkpoint that Google's next model may not escape, and why China is the open exception — a field reading of the new rules of access to frontier AI, for SMEs.

  • ai
  • ai-governance
  • geopolitics
  • frontier-models
  • europe
  • agi
Fable became mythic: when frontier models become strategic infrastructure

Fable became mythic

When frontier models stop being products and become strategic infrastructure

Difficulty: Intermediate — no code, no prerequisites. Just a bit of geopolitics and a healthy dose of suspicion.

In two weeks of June 2026, three things happened that, to me, matter far more together than apart.

On 12 June, Anthropic was forced to take two of its most capable models — Fable 5 and Mythos 5 — entirely offline, for everyone, after a U.S. government export-control order.

On 13 June, the Chinese lab Z.ai shipped GLM-5.2: an open-weight model you can download and run on your ‘own’ machines, with a one-million-token context window, for a fraction of the price of any Western frontier model.

And on 25 June — the day before I write this — the Trump administration asked OpenAI to release its next model, GPT-5.6, only to a handful of partners, with the government approving access customer by customer.

Three companies. Three stories. One direction.

Taken on their own, each of these has a tidy explanation:

* Anthropic: national security.
* OpenAI: a cautious rollout.
* Z.ai: Chinese acceleration.

Taken together, they say something less comfortable.

Access to the best models is no longer decided only by the companies that build them. Increasingly, it is decided by governments.

And the question quietly shifts from which model is best? to who is allowed to use it? — which, for anyone building on top of these tools, is a very different question.

Let me be clear about what I am not saying. Nobody has announced AGI. There is no verified proof that any lab is running true general intelligence behind closed doors. I am not claiming we have crossed that line. But something else is happening, and it may be almost as important:

the most powerful models are starting to be treated like strategic capabilities rather than products.

Once that shift happens, the technical league table stops being the only thing that matters.

1. The week access became a political decision

Start with Anthropic, because it is the clearest case.

Fable 5 and Mythos 5 launched on 9 June. Mythos 5 was the sensitive one, available only through a restricted government program called Project Glasswing, with some safeguards lifted and very strong cybersecurity capabilities. Fable 5 was the same underlying model made safe for broader use — a “Mythos-class” model with the dangerous capabilities walled off. Three days later, on 12 June at 5:21pm Eastern, Anthropic received a Commerce Department directive and pulled both.

The wording matters. The order did not target “non-US customers” or “companies abroad.” It targeted any foreign national, whether inside or outside the United States — including Anthropic’s own non-citizen employees. Since you cannot reliably check nationality at the API layer in real time, the company had only one practical option: disable the models for everyone. Other Claude models, including Opus 4.8, kept running.

The stated trigger was narrower than “this model is too powerful.” Officials told Anthropic they had learned of a technique to bypass Fable’s safeguards and reach Mythos’s cyber capabilities. Anthropic’s reply, in substance: the jailbreak is narrow, not universal, and the same trick works on other publicly available models — OpenAI’s GPT-5.5 included — that face no such restriction. The company called the whole thing a misunderstanding and said it was working to restore access. As of this writing, both models are still down. One former government AI-policy figure called the move “cartoonish.” Others said Anthropic was reaping what it sowed, having marketed Mythos as too dangerous to release in the first place.

You can argue the merits both ways. The operational lesson does not depend on who is right. A model can be launched, used, woven into real workflows, and become strategically important — and then vanish overnight, for reasons that have nothing to do with the model’s quality or your contract. For the customer, technical performance suddenly becomes the secondary question. The real one is: can I still access this tomorrow?

There is a small irony in the name. Fable 5 became unavailable a few days after launch. It became a thing people talked about, compared against, regretted. It became, almost literally, a myth. The joke writes itself, but the point underneath is serious: when a model disappears for geopolitical reasons, it stops being a normal product and becomes a signal. It tells every company outside the United States that access to frontier intelligence may depend on decisions taken outside their country, outside their contract, and outside their control. That is not an operational risk. It is a strategic dependency.

2. From emergency brake to pre-release checkpoint

The Anthropic case could still be read as exceptional — a sensitive model, a cyber concern, a government reacting after the fact. The OpenAI case is harder to wave away, because the intervention comes before release.

On 25 June, reporting from The Information, Reuters and Axios converged on the same story: OpenAI will ship GPT-5.6 first as a limited preview to a small set of partners, and during that window the government will be — in the words of a staff memo from Sam Altman — “approving access customer by customer during this preview period.” Axios added the texture: the White House had been briefed on the model’s capabilities, Altman discussed GPT-5.6 with Commerce Secretary Howard Lutnick, and the intervention was tied to the model having “Mythos-like” capability. A broader release is expected a couple of weeks later, if the approval process allows.

This is not the same pattern as Anthropic, and the difference cuts both ways. The OpenAI arrangement looks cooperative rather than coercive — a staggered rollout the company agreed to, not a directive that yanked a live product. But it is also, by several accounts, the first time the U.S. government has preemptively asked an American AI company to restrict a launch before release. The state is no longer only an emergency brake. It is becoming a pre-release checkpoint.

The framing makes this look softer than it is. Earlier in June, an executive order asked AI companies to voluntarily submit their most powerful models for review before shipping. Two weeks later, the government took Anthropic’s models offline and gated OpenAI’s. When the government asks and both labs comply, “voluntary” is doing less and less work. None of this means every model now needs a permit. It means the most capable frontier models may be entering a different category — capabilities whose release is coordinated, observed and approved, rather than simply pushed to production. And the framework is drifting toward a logic of capability rather than company: the voluntary reviews it sets up are aimed at the most powerful frontier models from any U.S. developer — even if, on paper, it still stops short of a mandatory permit for every powerful model. Google has not been through this yet — Gemini 3.5 Pro’s slip from June to July is, by every account, ordinary refinement — but the day it ships a model in the same class, there is no obvious reason it would escape the checkpoint that just caught its two rivals. For anyone who treats “API access” as a stable foundation, that is worth sitting with.

3. What “Mythos-like” actually means

The expression is doing a lot of quiet work, so it is worth unpacking. It does not mean GPT-5.6 is a copy of Mythos 5 — same weights, same architecture, same behaviour. The phrase is political, not technical. It means: this model may sit close to the capability band that recently spooked the government. Cybersecurity, autonomous action over long horizons, tool use, agentic workflows.

This is where the frontier gets hard to describe. For years we measured models with benchmarks — MMLU, GPQA, SWE-bench, Terminal-Bench, the rest. But the capability that now draws government attention is not a score on a static test. It is the ability to keep working: to plan, use tools, debug, adapt, persist, and chain steps over time to solve problems that used to need a human expert for hours or days. Once a system can do that reliably, the line between “assistant” and “operator” starts to blur. And an AI that can operate, rather than just answer, is operational power — which is the thing states have always paid attention to.

4. The signal might be silence

A quick, honest aside — and I will flag it as speculation, because that is the point. Could a lab cross something close to AGI internally before any of us hear about it? Possibly. Not proven, not knowable from the outside — but possible, for a dull structural reason: we only ever see the productized version. Labs see their best internal models first, governments sometimes see them next, and the public sees them last, if at all.

Mythos is the case in point. It was trained and red-teamed in early 2026, then announced on 7 April — but only to a few dozen vetted partners, for defensive security work. The public never got Mythos at all; we got Fable 5, the safeguarded version, two months later in June, and even that lasted three days before the government pulled it. The most capable model in the whole story was never something you or I could run.

So a real threshold might not arrive as a press release. It might arrive exactly the way this month did — a model delayed, an access window narrowed, a rollout restricted to approved partners, a launch reversed.

AGI may not arrive as an announcement. It may arrive as an access restriction.

And “AGI” is the wrong word to fixate on anyway: it means five different things to five people, from “can do any human task” to “superintelligence.” The question that actually moves governments is narrower and more useful — what can this model do that shifts the balance of power? Once the answer tips toward “possibly,” it gets controlled, whether or not anyone calls it AGI.

5. China isn’t waiting

While Washington tightens, Beijing ships. On 13 June, Z.ai launched GLM-5.2, with the open weights following under an MIT license a few days later: an advertised one-million-token context window usable on many workloads, and — depending on which benchmark you trust — coding performance in the neighbourhood of the Western frontier at roughly a sixth of the price. Treat the “beats GPT-5.5” headline as a vendor claim; the independently verified part is enough on its own. Chinese open-weight models are now good enough to move the market — not because the challenger wins everywhere, but because “good enough, cheaper, open and yours to run” beats “slightly better, expensive, closed and politically fragile” for a very large number of companies.

Here is the twist, and it matters to European buyers more than any benchmark:

A model’s nationality decides almost nothing about where your data lives — the host and the jurisdiction decide almost everything.

Z.ai’s own API raises a real question of jurisdiction, contract and effective data residence — one you settle at the provider and DPA level, not by reading the model’s flag. But GLM-5.2 is open-weight, so the identical model is also served by Western hosts (Fireworks, DeepInfra, Together AI) or your own air-gapped GPUs, none of which run through Z.ai’s own endpoint. Microsoft is about to make the same point at enterprise scale: it is exploring a fine-tuned DeepSeek V4 — or another open-source model — as a budget engine for its Copilot Cowork agent, hosted entirely on Azure so, the company says, the data never leaves its own cloud. So “open” does not automatically mean “sovereign,” and “Chinese model” does not mean “your data goes to Beijing.” What you check is who operates the endpoint and under whose law — with the weights’ origin a separate, provenance-level question that hosting does not erase. And the wider pattern is the one I described in my piece on AI and jobs: export controls do not stop China, they hand it a motive. When Washington choked off chips, Nvidia’s share of the Chinese AI-chip market it could still address fell from 95% to zero — by Jensen Huang’s own account — and Beijing turned the squeeze into a self-reliance program. One honest caveat: today’s Chinese openness is a strategy, not a law of nature. Beijing already controls which AI services reach its own public; it gives its weights away abroad because open release undercuts closed American models and builds the world’s dependence on Chinese ones — but the day a model looks too valuable to hand out, it could gate it exactly as Washington just did. For now, though, the contrast holds: Washington wants to protect its lead; Beijing wants to shed its dependency. Europe, for now, risks depending on both.

What it means for your SME

All of this can feel far away from a 30-person company in Belgium. GPT-5.6, Fable 5, GLM-5.2, export controls, sovereignty — it sounds like someone else’s problem. The practical consequence, though, is simple and immediate.

Do not build your business on the assumption that your favourite model will always be there. The best model today may not be accessible tomorrow, a provider’s roadmap is not independent of its government, and “API access” is not strategic control. This does not mean every SME should self-host a frontier model — that would be unrealistic. It means every serious AI project should be able to answer a short list of questions before it goes live. Which model, for which task, on which data? From which provider, under which jurisdiction? With which fallback, which evaluation, which logging, which human in the loop, and which exit plan if the model disappears?

These are not abstract governance questions. They are survival questions. If a key model goes dark, your workflow should degrade gracefully — slower, dumber, more expensive for a while — not collapse. Multi-model by default, fallbacks in place, sensitive workloads kept where you control them. It is less exciting than “we use the best model.” It is also what keeps you running on a Friday afternoon when a directive lands at 5:21pm.

And if you are European, hear the flip side. Europe will not win this by building “its own OpenAI” — with less capital and twenty-seven national strategies, that race only formalizes the delay. But the very things Europe is so often told are a handicap — auditability, compliance, control — are exactly what “trustworthy in production” demands. That is a real edge, and it is the most usable part of this whole story: you do not need the biggest model, you need the one you can run, trust and stand behind.

What I take from this

Three ideas I want to keep together.

Capability, access and sovereignty have become inseparable. For years companies treated AI like software — pick the best vendor, negotiate the price, integrate the API, measure the ROI, move on. That approach is no longer enough, because frontier models are turning into strategic infrastructure, and strategic infrastructure is never neutral. It is shaped by national security, export controls and the race between blocs.

The closer models get to AGI-like capability, the less likely they are to ship like normal products. That may be the most important signal of all — not the benchmark, not the demo, not the launch page, but the release pattern. Who gets access first? Who approves it? Who is excluded? Which models quietly disappear? Those are becoming the real indicators of where the frontier actually is.

For Europe, the answer is lucidity, not panic — and not regulation reduced to paperwork. Stop chasing a race written in San Francisco; compete where trust, control and domain knowledge decide the winner, not raw model size.

To wrap up

I do not know whether AGI has already been reached somewhere internally. Nobody outside a very small circle honestly can, and you should be wary of anyone who claims otherwise. But one thing is becoming clear: the world is starting to behave as if some AI models are no longer ordinary technology. Governments are watching, labs are coordinating, access is narrowing, open alternatives are accelerating. The U.S. is protecting its lead, China is reducing its dependency, Europe is trying to organize — and companies are caught in the middle.

Fable became mythic because it disappeared. GPT-5.6 may matter less for what it can do than for how access to it is being controlled. GLM-5.2 matters not only because it performs, but because it proves the alternatives will keep coming. And the AGI race, if it is happening, may not arrive with fireworks. It may arrive by silence — by delays, by restricted previews, by export controls, by customer-by-customer approvals, by models suddenly too powerful to release the normal way.

Use the best models when they make sense. But never forget that access is not ownership, and dependency is not a strategy. That is the world worth preparing for.

Until the next one — and keep asking the only question that counts: who decides whether I still get to use my AI model tomorrow?

Sources (checked 26 June 2026; vendor benchmark claims and unverified reporting are flagged as such):

All articles
  • ai
  • ai-governance
  • geopolitics
  • frontier-models

Ready to go further?

Let's discuss your case in 20 minutes. Free, concrete, no strings attached.

Book a slot