Anthropic unplugged: how the shutdown of Fable and Mythos changes your AI strategy

On July 1, 2026, the US government forced Anthropic to pull its two most powerful models — Fable 5 and Mythos — from general circulation. Official reason: an export control directive banning access by any non-US national anywhere on the planet.

Unable to filter users by nationality, Anthropic shut everything down for everyone — including its US customers and its own non-US engineers.

What happened, hour by hour

The directive arrived at 11 PM with zero detail on the concrete threat. The alleged breach? Users jailbroke Fable 5 using known techniques — role-play reformulation, base 64 obfuscation, constitutional classifier bypass.

According to the Wall Street Journal, Amazon reported the flaw to the Department of Commerce. Amazon is Anthropic s largest investor ($33B) but also the Pentagon s network provider. To protect its military contract, the giant turned in its own protégé.

The real reason: a 6-month vendetta

The security flaw is a pretext. The story begins in February 2026, when the Pentagon demanded unrestricted Claude use for all lawful purposes — including autonomous weapons and mass surveillance.

Anthropic refused publicly. The retaliation was immediate. Trump ordered all federal agencies to cut Anthropic. A federal judge in San Francisco ruled this designation was a pretext for unlawful retaliation against a company exercising free speech.

In May, the Pentagon signed classified network deals with Palantir and Anduril. Not Anthropic. The July directive follows in this wake.

The 3 walls of AI control

This affair reveals a control architecture built in three layers:

Wall 1 — Chips (since 2022)

Hardware stops at customs. No NVIDIA H100s to China. No ASML lithography machines. TSMC fabricates under US supervision. Physical, countable, controllable.

Wall 2 — Money (Fable 5, 2026)

Frontier models are so expensive they naturally filter by wallet. Fable 5 runs at roughly $100/million tokens — out of reach for SMBs.

Wall 3 — Weights (July 2026) — THE TURNING POINT

For the first time, a government judges a model at source and decrees an object too dangerous for certain hands. Your passport becomes the access criterion. Unlike the first two walls, this one cannot be crossed by paying.

What this changes for European companies

1. Single-vendor dependency is a death sentence

If your main provider cut you off tomorrow morning, how long before your clients notice?

2. Chinese open models are now credible alternatives

Top open models today are Chinese: DeepSeek, Qwen, Kimi, Yi. They rank within points of closed US models. Free, downloadable, commercially licensed.

3. Mistral gains strategic meaning

The only serious open model that is neither US nor Chinese. Behind on bleeding-edge benchmarks, but entirely European.

4. Data residency becomes a sales argument

Host at OVHcloud or Scaleway instead of US-jurisdiction cloud. Yesterday a compliance burden, tomorrow a competitive advantage.

The irony

Anthropic built its brand on its models danger. Sam Altman mocked it: "Genius marketing to say you built a bomb, then sell the shelter for $100 million."

The joke backfired. The government took Anthropic at its word. When you describe your product as munition in every press release, you write the law that bans it.

The loop closes

A closed model keeps its guardrails on the provider s servers. An open model you host yourself works in reverse: you hold the weights, you control the guardrails. And those can be removed.

The danger? Any motivated developer can take an unaligned model and use it to mass-produce jailbreak prompts against other models. The defender needs perfection. The attacker needs one breach.

The takeaway

We are leaving the era where strategy meant plugging in the best US model and building on top. Comfortable and fragile.

The coming world rewards compute sovereignty, open-model mastery, and data residency. And the increasingly rare profile of someone who knows, for each task, which model runs where and under which jurisdiction.

This week s concrete action: Map your critical processes. Next to each one, write which model runs it and under which jurisdiction. Every line that says "US frontier model" is a switch someone else holds in Washington.

You don t have to change everything tomorrow. But you will no longer decide blind.

Go further on multi-model strategies and agentic orchestration: AI French Touch 1000 Skills Hub.