8SEP

US order pulls Anthropic's top models

3 min read

15:06UTC

The Trump administration ordered Anthropic to bar foreign nationals from Fable 5 and Mythos 5; unable to screen by nationality, the company pulled both models worldwide.

← Back to Mistral closes EUR 3bn; Luxembourg buys in Jump to analysis ↓

TechnologyDeveloping

Key takeaway

A single US order pulled two frontier AI models from every European user at once.

The Trump administration ordered Anthropic to block foreign nationals from its two most capable models, Fable 5 and Mythos 5, on national-security grounds. Screening every user by nationality proved technically unworkable, so Anthropic took both models offline for everyone rather than enforce the screen. Axios first reported the order on 12 June, and CNBC, Fortune and Anthropic itself later confirmed it.

Christophe Grudler, a French member of the European Parliament, called the move proof that the US holds "a real kill-switch over essential technologies". A European Commission spokesperson said firms must comply with EU legislation whatever their home government demands. The cut reached inside the bloc's own institutions: the EU Agency for Cybersecurity (ENISA) had been admitted to Anthropic's Project Glasswing government-access programme in April, then lost access to the relevant models within weeks of joining.

Bruegel had already documented that Europe depends structurally on US or Chinese compute despite its sovereignty drive . A single US order cutting off two frontier models for every European user at once is the first working instance of the dependency that analysis described. European enterprises now know the capability they license from an American lab can be revoked by an American agency, not the vendor, and the design of the product cannot stop it.

Deep Analysis

In plain English

Anthropic is an American artificial intelligence company. Its most capable AI models, Fable 5 and Mythos 5, were being used by European businesses, researchers and government agencies. On 12 June, the US government ordered Anthropic to stop letting foreign nationals use these models for national security reasons. The problem was that checking every user's nationality turned out to be technically impossible, so Anthropic switched the models off for all users globally, including Europeans and EU institutions. That included ENISA, the EU's own cybersecurity agency, which had been given special government-level access just two months earlier. The episode showed that access to some of the most powerful AI tools available can be cut off by a single US government decision, with no warning and no EU veto.

Deep Analysis

Root Causes

Anthropic's corporate structure gives the US government a jurisdictional hook that European AI companies lack.

Incorporated in Delaware and receiving US Department of Energy compute allocations, Anthropic is subject to export administration regulations that allow the Commerce Department to restrict model access on national-security grounds without judicial review, a control that does not apply to a French-incorporated Mistral or a German-incorporated Aleph Alpha operating the same model weights on European infrastructure.

The EU's own CADA (Cloud and AI Development Act), adopted on 3 June, mandates that public bodies route sensitive workloads through EU-controlled infrastructure, but left frontier model access unresolved. CADA's public-sector cloud tier was designed around data storage and processing, not model inference, meaning ENISA could route its compute through an EU cloud while the model weights themselves remained under US export jurisdiction.

Escalation

Direction is toward further restriction, not relaxation: the Commerce Department's 27 June partial reversal kept conditions in place rather than fully restoring access, and the 2 August GPAI enforcement deadline gives Brussels a regulatory hook to demand model continuity guarantees that US providers operating under US export control cannot currently give.

What could happen next?

Consequence
European enterprises building regulated AI workflows on US frontier models now have a documented precedent for unilateral cut-off, accelerating the business case for migrating to EU-incorporated alternatives such as Mistral and Aleph Alpha.
Short term · Assessed
Risk
From 2 August, GPAI-regulated enterprises in the EU that relied on Anthropic models could face compliance gaps if a future US order cuts access without the 15-day reversal cycle seen in June.
Short term · Reported
Precedent
The episode establishes that US export controls on AI models operate under the same legal authority as munitions controls, making future restrictions on allied-country AI access legally routine rather than exceptional.
Long term · Assessed

Source Landscape

This story draws on centre-left-leaning sources from United States

United States×2

Primary parallel: In 1996, the US Commerce Department classified strong encryption software as a munition under ITAR, barring export of PGP and similar tools to foreign nationals without a licence.

The policy proved unworkable: source code was printed as a book and exported legally, and the restriction was largely lifted by 2000. Anthropic's nationality-screening failure in June 2026 is structurally identical: algorithmic capability, like cryptographic code, cannot be screened by passport at the access layer.

Counter-parallel: The Wassenaar Arrangement, adopted in 1996 by 42 states to coordinate dual-use technology export controls, successfully restricted EUV lithography machine exports to China for over a decade without a unilateral US override of Dutch licensing. The arrangement worked because it was multilateral and prospective; the Anthropic order failed because it was unilateral and retroactive to existing access relationships.

Consensus view: Andrea Renda at the Centre for European Policy Studies (CEPS) assessed in May 2026 that the EU's frontier-model dependency on US providers created an 'asymmetric compliance obligation': EU enterprises must comply with GPAI obligations from 2 August while the model providers they depend on can be pulled by a unilateral US government order, leaving European buyers in regulatory breach through no action of their own.

Counter-view: The Observer Research Foundation's Bedavyasa Mohanty, writing from New Delhi, argues that Washington's order reflects a defensible national-security calculus: frontier models trained on sensitive data carry export-control logic equivalent to munitions, and allied states accepted similar constraints under ITAR (the International Traffic in Arms Regulations) for physical technology throughout the Cold War.

Key tension: Whether EU enterprise buyers facing the 2 August GPAI compliance deadline can legally demonstrate 'adequate measures' toward regulatory compliance when their primary frontier model provider can be switched off by a foreign government on 15 days' notice.

First Reported In

Update #10 · Digital euro to trilogue; Senate bars CBDC

EU Perspectives· 30 Jun 2026

Read original →

Causes and effects

This Event

US order pulls Anthropic's top models

A US national-security order, not the vendor, removed Europe's access to two frontier models overnight, turning bought AI capability into a revocable licence.

Led to

US partly lifts its Anthropic model curb

The original Anthropic worldwide pulldown is the precipitating event that the US Commerce Department's partial lift on 27 June partially reversed.

Occurred 27 Jun 2026

Read story →

MPs name the UK's sovereignty gap

The Commons committee report names the Trump administration's order barring foreign access to Anthropic's models as its trigger case for the sovereignty-gap finding.

Occurred 7 Jul 2026

Read story →

Different Perspectives

Poland

Poland leads a self-announced AI Gigafactory consortium with a EUR 100 million phase-one commitment, matched by Czechia and joined by Hungary at EUR 25 million. EuroHPC has confirmed no consortium for the call closing 12 November, so the bloc exists only in national announcements so far.

UK Government (DCMS)

Secretary of State Lisa Nandy told the Commons on 3 September she has inherited the sovereign AI brief from Liz Kendall, but gave no assessment, figure or date on frontier-model access. It is the first public claim of ownership since DSIT's abolition, without the substance the committee asked for.

ASML

ASML CEO Christophe Fouquet credited Intel Foundry, not a European fab, with shipping the first high-volume logic product made on High-NA EUV, the tool only ASML builds. Europe holds the chokepoint tool; the company that spent it into volume production first is American.

Luxembourg

Luxembourg joined Mistral's Series D as a new investor on 8 September, the same government that co-funds EuroHPC's MeluXina-AI supercomputer. One small member state now funds the sovereign compute Mistral may need and holds equity in the company using it.

Samsung Electronics

Samsung led Mistral's Series D on 8 September and the same day expanded its ASML collaboration on next-generation lithography. One Korean company now sits atop Europe's largest AI funding round and inside its most sensitive chip-tooling relationship.

Mistral AI

Mistral closed a EUR 3 billion Series D above a EUR 21 billion valuation on 8 September, with Samsung leading and Luxembourg joining as a new state investor. The company markets itself as Europe's non-American AI option even as the round's lead capital comes from South Korea and two US private-equity firms.