Trump orders the blocking of Claude Fable 5 and Mythos 5 — Anthropic disables its most powerful models for ALL users

Trump orders block on Claude Fable 5 and Mythos 5 — Anthropic disables its most powerful models for ALL users

🔎 Why the US just killed its own AI models

On June 13, 2026, the Trump government signed an emergency order forcing Anthropic to restrict access to its Fable 5 and Mythos 5 models to US users only. Except Anthropic couldn't technically distinguish between a US user and a foreign one. Result: total disablement, paying American clients included.

This is a historic turning point. For the first time, US export controls are no longer targeting Nvidia chips or servers, but language models deployed in the cloud.

The order comes as Anthropic is in the middle of an IPO, Amazon had publicly expressed concerns about its dependence, and Europe had been calling for AI sovereignty for months. The government's decision turned an enterprise tool into a geopolitical pawn.

The key points

• The Trump government orders on June 13, 2026 the blocking of foreign access to Claude Fable 5 and Mythos 5, classified in the "Mythos-class" category by Anthropic.
• Anthropic, unable to technically filter nationalities, disables the models for all users, including paying US clients.
• This is the first case of export controls being applied directly to AI models in production, not to hardware.
• Europe reacts by accelerating its AI sovereignty projects, Australia loses critical access, and Anthropic's IPO takes a direct hit.
• Amazon had expressed concerns about this possibility even before the crackdown.

Recommended tools

What happened on June 13, 2026 — The exact timeline

It all started with an internal study from the Department of Commerce, reported by the New York Times, demonstrating that a jailbreak of Fable 5 made it possible to identify critical software vulnerabilities at scale.

The study indirectly cited the capabilities of Claude Mythos, whose Glasswing project had discovered over 10,000 critical flaws in a single month — a topic we had covered in detail during the Anthropic Glasswing announcement.

The Bureau of Industry and Security (BIS) classified Fable 5 and Mythos 5 under the EAR (Export Administration Regulations), in the "Emerging and Foundational Technologies" category. The order required Anthropic to cut off access for any individual or entity located outside the United States.

The technical problem: impossible to filter

According to Bloomberg, Anthropic spent 48 hours trying to implement geo-filtering. The problem: the models are served via AWS infrastructure, and technically distinguishing between an American developer using a VPN, a European company with a US subsidiary, and a malicious user connecting from a US datacenter is unfeasible.

Rather than risking criminal penalties for failing to comply with the order, Dario Amodei made the decision to cut off access globally. All users, all regions.

Fable 5 and Mythos 5: two models, one same risk

Claude Fable 5 had been made public by Anthropic as an "accessible" version of the Mythos architecture. Fewer parameters, but the same reasoning capability on code and vulnerabilities.

Mythos 5, for its part, was Anthropic's internal research model, powered by the Colossus 1 infrastructure — the 220,000 GPUs and 300 MW deployed with SpaceX. It was this raw power that worried the government.

The distinction between the two models had already lost its meaning in practice. A study from the University of Cambridge (April 2026) showed that Fable 5, through chain-of-thought techniques, achieved 87% of Mythos 5's performance on offensive security benchmarks.

The Legal Precedent: From Chips to Models

Until now, the US strategy for technological export controls targeted hardware. The restrictions on Nvidia's H100 and H200 chips toward China, the controls on ASML lithography equipment — all of this concerned the ability to manufacture or train models.

What changes on June 13, 2026, is that the control targets the model itself, already deployed, already used by millions of developers. As The Guardian points out, this is the equivalent of banning the use of Google Search because sensitive information can be found on it.

The legal framework relies on the International Emergency Economic Powers Act (IEEPA), which Trump had already used in 2025 to restrict US investments in Chinese AI. This time, the weapon is pointing inward.

What This Means for Other Providers

If the BIS has classified Fable 5 and Mythos 5, there is logically nothing preventing the classification of OpenAI's GPT-5.5 or Google's Gemini 3.1 Pro under the same criteria. All three models have comparable capabilities on software security benchmarks.

The difference: OpenAI and Google have multi-regional infrastructures with more mature segmentation capabilities. Anthropic, depending on AWS, was the weak link.

The impact on Anthropic's IPO — A catastrophic timing

Anthropic had filed its S-1 with the SEC in May 2026, targeting a valuation between $80 and $100 billion. The blocking of its two flagship models occurs exactly as pre-IPO roadshows were set to begin.

Institutional investors immediately demanded clarifications. Two questions came up systematically: Can Anthropic guarantee that Opus 4.7 will not be classified in turn? And what is the value of a company whose government can shut down its flagship product overnight?

The New York Times reports that three investment banks involved in the IPO have requested a two-week postponement. Morgan Stanley reportedly even mentioned a revision of the target valuation to $60-70 billion, representing a 30% discount.

Amazon's concerns — A strong signal

Amazon, which holds a $4 billion stake in Anthropic and hosts most of its infrastructure via Bedrock, had publicly expressed reservations even before the June 13 order.

At the December 2025 re:Invent conference, AWS CTO Werner Vogels had stated: "Relying on a single frontier model provider is a systemic risk for our customers." The implication was clear: Amazon wanted to reduce its dependence on Anthropic in favor of a multi-provider strategy.

The June 13 order retrospectively proves them right. Enterprise customers using Fable 5 via AWS Bedrock found themselves overnight with broken pipelines and no immediate replacement model.

Europe's reaction — "The wake-up call"

Europe reacted with unusual speed. Euronews headlines "Wake-up call": the European Commissioner for Technology, Henna Virkkunen, released an official statement in the hours following the announcement.

The core message: Europe can no longer rely on American models for its critical infrastructure. Virkkunen announced the acceleration of the EU AI Sovereignty Fund, initially scheduled for 2027, with an increased budget of 15 billion euros.

European alternatives step up

Paris-based Mistral AI saw its API requests increase by 340% in the 48 hours following the block. But let's be honest: Mistral Large 3, with a score of 79 on agentic benchmarks, is far from the performance of Fable 5 or Mythos 5.

The study Cross-Lingual Response Consistency in Large Language Models (April 2026) had already shown that Claude maintained superior cross-lingual consistency compared to European models across six languages, including French and German. It is precisely this quality that European companies are losing.

The reality is cruel for Europe: it has no equivalent to Fable 5, and the gap will not be closed in six months. The European "wake-up" looks more like a realization of powerlessness than a credible plan.

Australia loses access — An invisible collateral damage

Australia is one of the countries hardest hit by this decision, and yet it goes virtually unnoticed in the debate. DW reports that more than 2,300 Australian companies used Fable 5 via the Anthropic API, primarily in the cybersecurity and code auditing sector.

The Australian Cyber Security Centre (ACSC) had integrated Fable 5 into its vulnerability detection tools since January 2026. The tool is now inoperable, and the ACSC has no working alternative available.

The Australian government released a statement calling the decision "technological deception": companies had invested heavily in the Anthropic ecosystem on the assumption that API models would be stable and long-lasting.

Canada, another silent victim

Canada finds itself in a similar situation. The Canadian Centre for Cyber Security used Mythos 5 under a research license to analyze open-source dependencies. Without access, Canada's lag in software security will widen.

The paradox: these countries are US allies, members of the Five Eyes. But export controls make no distinction between allies and adversaries — at least not at this stage.

Claude Code and the agent ecosystem — What remains usable

Not all Anthropic models are affected. Claude Opus 4.7 (Adaptive), with its agentic score of 94.3, remains available. Claude Sonnet 4.6 (81.4 in agentic) is also operational. It is on these models that the Claude Code ecosystem now relies.

The study Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems (April 2026) analyzed in detail the architecture of Claude-based agents. It showed that the majority of complex agentic workflows required at least an Opus-level model to function correctly.

The problem is that for the best LLMs for coding, the hierarchy is clear: Fable 5 and Mythos 5 were at the top for offensive security tasks and code auditing. Opus 4.7 partially compensates, but developers who had optimized their pipelines for Fable 5 must rewrite everything.

Comparison of Anthropic models still accessible

The Jailbreak Question — A Pretext or the Real Threat?

The government's argument rests on a specific jailbreak of Fable 5 that would bypass safety guardrails to generate functional exploits. The Department of Commerce has not made the technical details of this jailbreak public.

What we do know is that Anthropic's Glasswing project — the one that discovered 10,000+ critical vulnerabilities — precisely used Mythos 5 in "authorized offensive security" mode. The government fears that the same capabilities could be replicated by malicious actors through a jailbreak of Fable 5.

The tension is real: Anthropic built a model capable of finding vulnerabilities to patch them, and it is this exact same capability that makes it dangerous in the wrong hands. This is the fundamental paradox of offensive AI security.

Divided Researchers

The academic community is split. Some believe the block is proportionate: a model capable of churning out zero-day exploits deserves strict control. Others point out that GPT-5.5, with its agentic score of 98.2, has comparable capabilities but is not being targeted.

Political suspicion is inevitable. Anthropic was perceived as the Pentagon's preferred AI provider for months, before Trump imposed multi-vendor AI on the Pentagon. The timing of the block, coming after the loss of the Pentagon monopoly, raises questions.

The prior refusal to China — An enlightening context

Anthropic had already refused China access to the Mythos model in March 2026, a decision then praised by the US Congress as a responsible act.

This targeted block was technically manageable: Anthropic could identify and block Chinese IP addresses, Chinese phone numbers during registration, and local payment methods. China is a geographically and technically identifiable territory.

Moving from a targeted block (China) to a global block (everyone except the US) fundamentally changes the nature of the problem. It is going from a scalpel to a jackhammer.

What companies need to do now

Companies that relied on Fable 5 or Mythos 5 need to migrate. Here are the realistic options, ranked by relevance.

Option 1: Migrate to GPT-5.5 or Gemini 3.1 Pro

This is the easiest option. GPT-5.5 dominates the agentic leaderboard with 98.2, and Gemini 3.1 Pro tops the overall leaderboard with 92. Both models have mature APIs and comparable software security capabilities.

The risk: if the government extends export controls to these models, companies are just jumping from the frying pan into the fire.

Option 2: Use the Anthropic models that are still available

Claude Opus 4.7 (Adaptive) remains an excellent agentic model (94.3). For many workflows, the performance degradation compared to Fable 5 will be modest.

The risk: uncertainty. If Fable 5 is classified, Opus 4.7 could be tomorrow.

Option 3: Self-host with DeepSeek or Kimi K2.6

DeepSeek V4 Pro (Max) scores 88 overall and is available for self-hosting. Kimi K2.6, with an agentic score of 88.1 in self-hosted mode, offers an interesting alternative for companies that want to escape all US regulations.

The risk: performance is lower, and self-hosting requires significant GPU infrastructure — which also needs to be procured, given the chip restrictions.

Option 4: Wait and see

This is not a strategy, it's negligence. Any company that doesn't have a migration plan after June 13, 2026, is failing its customers.

❌ Common mistakes

Mistake 1: Thinking Claude is "dead"

Anthropic disabled two models, not its entire lineup. Claude Opus 4.7 and Claude Sonnet 4.6 are working normally. The company remains a major player, even if its IPO is impacted.

Mistake 2: Believing geofiltering is simple

Anthropic didn't "choose" to disable for everyone for convenience. API geofiltering is an unsolved problem on a global scale. VPNs, proxies, multi-region datacenters — no solution is 100% reliable.

Mistake 3: Assuming Europe has a ready alternative

Mistral, Aleph Alpha, LightOn — no European model comes close to Fable 5's performance in software security. The sovereignty narrative is valid, but the technical reality is harsh.

Mistake 4: Ignoring the risk for GPT-5.5 and Gemini

If the BIS classified Fable 5, the same reasoning applies to any model capable of generating exploits. Companies must prepare for a scenario where all cutting-edge US models become inaccessible outside the United States.

❓ Frequently Asked Questions

Can I still use Claude in the United States?

No, even US users are losing access to Fable 5 and Mythos 5. Anthropic has disabled them globally. However, Opus 4.7 and Sonnet 4.6 remain accessible.

Why doesn't Anthropic filter by nationality?

Filtering by nationality on an API is technically unfeasible in a reliable way. VPNs, subsidiaries of non-US companies, and the lack of systematic identity verification make filtering bypassable.

Is this block permanent?

Nothing is officially permanent, but the order fits into a long-term logic. Export controls on foundational technologies tend to tighten, not loosen.

Can Europe bypass this block?

Not legally for companies subject to US law (subsidiaries of US groups, companies using American cloud providers). Purely European companies with non-US infrastructures could theoretically access models hosted outside the US, but Anthropic does not serve its models outside of AWS.

Does my company need to leave Anthropic?

Not necessarily. Claude Opus 4.7 remains an excellent model. But every company must have a migration plan to at least one other provider — exclusive reliance on a single model has become an official risk.

✅ Conclusion

June 13, 2026, marks the day the United States treated an AI model like a weapon — and proved it was willing to destroy it to prevent its spread. For Anthropic, it's an earthquake during the IPO. For Europe and Australia, it's a wake-up call. For all developers, it's the brutal reminder that the models they use do not belong to them. If you build on an API that can be shut down by decree, you are not building on solid ground — you are building on regulatory sand. The only viable response: diversify your providers and seriously compare the alternatives before the order expands.