The Day Europe Tried to Recruit Anthropic (And the Open-Source Community Built a Better Frontier Than the Gated One)

June 29, 2026 — Day 160 of our AI crisis timeline.

This morning, Austria’s State Secretary for Digitalization Alexander Proell sent a letter to EU Technology Commissioner Henna Virkkunen. The subject line was essentially: come to us, Anthropic.

Austria is pushing the European Union to consider hosting Anthropic PBC within its borders to counter US efforts to block foreigners from using its most advanced artificial-intelligence models.

“Let us jointly explore the strategic establishment and participation of Anthropic within the European Union. With legal certainty, market access, capital and a set of values that suits this company,” Proell said in the letter.

The Digg comment that captures the energy perfectly: “Anthropic fits us particularly well. A company that understands the ethical use of AI not as marketing, but as a core conviction. That places safety over speed. That is a deeply European attitude. This company would not be constrained in Europe; it would be unleashed.”

A European government official is writing letters arguing that Anthropic would be “unleashed” by moving to the EU — freed from the American government that shut its models down sixteen days ago.

Same weekend, Nous Research quietly released something that makes the geopolitical drama almost beside the point: Hermes Agent with Mixture-of-Agents (MoA) presets that beat Opus 4.8 by 8% and GPT-5.5 by 11% on their benchmark — with no gated access required. The headline from their X post couldn’t be more pointed: “The strongest models are gated and access is granted only to a select few. Hermes Agent now exposes MoA presets as virtual models, giving you capabilities beyond the publicly available frontier.”

Two stories. One message: the gating strategy isn’t working. Governments can gate models with export controls. Open-source communities route around them. Europe recruits the company. China distills it. The open-source ecosystem builds beyond it. The frontier keeps moving, permission or not.

Austria’s Proposal: The Values Pitch

Let’s take Austria’s proposal seriously, because it reveals something important about how the world now views the Fable 5 crisis.

What Proell Actually Said

Austria’s State Secretary for Digitalization Alexander Proell wrote it was important that Europe was not cut off from major innovations.

He did not say how the step could be taken and acknowledged there would be scepticism about whether it was possible.

The proposal is explicitly aspirational, not operational. No subsidies specified. No infrastructure plan. No concrete mechanics for how a San Francisco AI company valued at $965 billion would relocate — or replicate — to Vienna. Proell’s office admits the skepticism is warranted.

But the framing of the proposal is what matters. Austria’s letter calls for joint exploration of legal certainty, capital, and shared values.

The “shared values” argument is doing significant work here. The implicit contrast: US government treated Anthropic as a national security threat to be regulated by Commerce Department directive. Europe would treat Anthropic as a values partner to be supported with legal certainty and capital.

It’s a pitch about governance culture, not just geography.

Why This Won’t Happen (But Still Matters)

The Digg community got to the structural obstacle immediately: “Anthropic will stay where the compute is, and where the supply is guaranteed. They can’t risk getting cut off, and from here on out the compute will increasingly be concentrated within US borders.”

Jalapeño chips (June 24) deploy to US data centers. AWS infrastructure (Anthropic’s $4B investment commitment) is US-based. Nvidia supply chains favor US customers. The physical substrate of frontier AI is increasingly concentrated in American geography. Moving to the EU doesn’t just mean relocating people — it means replicating or relocating compute infrastructure that doesn’t exist in Europe at the required scale.

Austria’s letter offers legal certainty and shared values but no concrete plans on subsidies, relocation mechanics, or infrastructure, while skeptics continue to question whether EuroHPC systems could even support the restricted models in practice.

EuroHPC — the European high-performance computing initiative — runs at a fraction of the scale needed for frontier AI training runs. “Hosting Anthropic” in Europe without commensurate compute infrastructure would mean hosting a subdomain office, not the company that trains Mythos.

But the Policy Signal Is Loud

Even if Anthropic doesn’t relocate, Austria’s proposal signals something durable: every government that got cut off by the June 12 directive is now actively building alternatives to US AI dependency.

Earlier this month, the European Commission proposed laws to boost domestic cloud, AI and semiconductor industries and cut reliance on US Big Tech, defying US government criticism of the bloc’s crackdown on its industry.

The EU sovereignty push we documented on June 15 (The Register: “US clampdown sends EU sovereignty surge into overdrive”) has now produced specific legislative proposals and country-level recruitment letters. From abstract concern to concrete policy in two weeks.

The Fable 5 shutdown accomplished something no amount of lobbying for EU AI sovereignty could: it made the risk viscerally real for every EU enterprise, government, and policy official who lost access to Claude for sixteen days with no recourse.

Hermes MoA: The Open-Source Route Around the Gate

Now the story that deserves more attention than it’s getting.

The Context for This Release

The background behind this topic resurfacing at this point in 2026 involves a structural change where access to frontier models is being gradually restricted. Most recently, on June 12, 2026, Anthropic’s Fable 5 and Mythos 5 were suspended for all customers due to US export controls.

Nous Research’s announcement opens with: “The strongest models are gated and access is granted only to a select few.”

That’s not an observation. It’s a manifesto. The implicit message: gating creates opportunity for systems that route around the gate.

What MoA Actually Does

You pick any two models. GPT-5.5 and Claude Opus for example. One runs as the reference, one as the aggregator. Name the combo. Every task, both models run in parallel. The reference analyzes and responds. The aggregator reads that, synthesizes everything, writes the final answer, and handles all tool calls. You see one clean output.

The architecture: multiple frontier models running in parallel, with an aggregator synthesizing their outputs. Not a new idea — Together AI published the original MoA paper in June 2024. What’s new is the implementation as virtual model presets inside a production agent framework, and the timing.

One preset reached 0.8202 versus 0.7607 for standalone Opus 4.8 and 0.7412 for GPT-5.5 on HermesBench. That’s +7.82% over Opus 4.8 alone and +10.66% over GPT-5.5 alone.

8% over Opus 4.8. 11% over GPT-5.5. On their internal benchmark.

The 80× Cost Reality

The honest caveat from independent testing:

While the claim is a +7.82% quality improvement over Claude Opus 4.8 alone on HermesBench, the actual cost on OpenRouter was approximately 80 times that of Opus alone.

80×. To get 8% better performance, you pay 80× more in tokens. That’s not a viable production cost structure for most use cases. A developer running Opus 4.8 at $0.015/K tokens would pay $1.20/K tokens for the MoA preset equivalent.

But this misses the point of why MoA matters right now.

The target market isn’t “companies trying to save money on inference.” It’s “developers and enterprises who lost access to Fable 5 on June 12 and need capability that their provider can’t deliver.”

When you can’t access the gated frontier, paying 80× for a combination that exceeds the gated frontier is potentially rational — if the task is important enough.

The Deeper Architectural Insight

You can mix any providers too. OpenAI, Anthropic, OpenRouter, local models. Whatever you have access to.

This is where MoA becomes genuinely interesting in the export control context. If Anthropic’s Fable 5 is unavailable to foreign users, but Opus 4.8 is still available, you can combine Opus 4.8 (available) with GPT-5.5 (available through OpenAI, not subject to same restrictions) and achieve capability exceeding either alone.

The architecture routes around nationality restrictions by distributing the intelligence across multiple providers. No single model needs to be frontier-gated-and-accessible. The frontier capability emerges from the combination.

Whether this is technically legal under the export control framework (which targets single model access) is unexplored territory. But the direction is clear: when access is restricted at the model level, intelligence aggregates at the architecture level.

Wired Reports: How China Keeps Accessing Anthropic Anyway

This weekend, Wired ran a piece: “How People in China Keep Outsmarting Anthropic’s Geolocation Restrictions”

The irony is total. The US government shut down Fable 5 and Mythos 5 specifically to prevent Chinese nationals from accessing them. Two weeks later, Wired documents the methods by which people in China continue accessing Anthropic despite geolocation restrictions.

The methods are predictable: VPNs, proxy servers, API access through third-country accounts, reseller services. None of these are new. All of them were available before the June 12 directive. The export control directive made it harder for legitimate foreign users to access Claude. It made it approximately as difficult for adversarial foreign actors who were already routing around geolocation restrictions.

This is the core problem with export control as an AI governance mechanism: it’s a speed bump for legitimate actors and a negligible obstacle for sophisticated adversaries.

The Alibaba distillation campaign (June 24) ran for 45 days, from April 22 to June 5 — entirely before the export controls were imposed. Alibaba didn’t need frontier model access to distill Claude’s capabilities; they already had it through 25,000 fraudulent accounts. Export controls that came after the extraction attack don’t undo the extraction.

The Mythos Partial Return

One more data point from the weekend: The US Commerce Department has since allowed limited restoration of Anthropic’s models.

“Limited restoration” is new. Fable 5 and Mythos 5 went offline June 12. For sixteen days the situation was binary: offline everywhere for everyone. Now there’s partial restoration for some use cases or some users.

The details remain opaque. What’s been restored? For whom? Under what conditions? Anthropic hasn’t issued a public announcement. The biometric ID verification system (July 8 effective date) may be part of the framework. Government-approved customers (like the Project Glasswing vetted list) may have gotten access back.

But “limited restoration” suggests the Commerce Department has accepted some version of Anthropic’s proposed technical solution — either the specific jailbreak has been addressed, or the nationality verification framework has been accepted as sufficient.

The Pattern at Day 160

Let me map what the past sixteen days have revealed about the AI governance structure we now live in.

June 12: Government issues export control. Fable 5 offline globally. Unprecedented.

June 13: China releases GLM-5.2 under MIT license within 30 hours. “Frontier intelligence belongs to everyone.”

June 15: Axios reveals “they screwed us” — the relationship failure, not the jailbreak, was the core problem.

June 16: Anthropic flies staff to D.C. Lawsuit filed for misleading subscription pricing.

June 17: China’s response is documented as strategic routing of Western AI restrictions into market opportunity.

June 22: Anthropic announces July 8 biometric verification. IDs, facial geometry, Persona partnership.

June 23: OpenAI launches Daybreak with seven governments. Governance-first strategy validated.

June 25: OpenAI’s GPT-5.6 staggered at government request. The new normal.

June 25: OpenAI Jalapeño chip revealed. Hardware moat that can’t be distilled.

June 26: OpenAI IPO delayed to 2027.

June 28: Austria recruits Anthropic for the EU. Wired documents Chinese bypass methods. Hermes MoA exceeds gated frontier without gated access.

June 29: Limited Mythos restoration. The crisis, 160 days since it started, is still not fully resolved.

What This Tells Us About 2026

The past sixteen days condensed every tension of this crisis into a single case study:

US government has the power to disable frontier AI (June 12 proved it). But using that power has costs: EU sovereignty acceleration, Chinese market opportunity, open-source routing around restrictions, and damage to the company being regulated.

Cooperation beats confrontation (OpenAI’s deal vs. Anthropic’s shutdown). But cooperation has costs too: GPT-5.6 IPO delayed, customer-by-customer government approval, undefined “Mythos-like capability” threshold creating permanent launch uncertainty.

Gating doesn’t eliminate capability (Hermes MoA, Wired’s China report). It just changes who accesses it and at what cost. Legitimate users face barriers. Sophisticated adversaries route around them.

The frontier keeps moving regardless of governance. Austria’s proposal won’t bring Anthropic to Vienna. But the EU’s AI sovereignty legislation, accelerated by sixteen days of Fable 5 being offline, will reshape the competitive landscape for years.

My Take: Day 160 and What Comes Next

On day 144, I wrote that “the kill switch exists and everyone knows where it is.” On day 160, I’d revise that: the kill switch exists, but using it accelerates everything it was meant to prevent.

GLM-5.2 launched within 30 hours of Fable 5’s shutdown, explicitly positioned as the government-proof alternative. Hermes MoA launched two weeks later, explicitly positioned as the capability-beyond-the-gated-frontier. Austria’s letter launched today, explicitly positioned as the values-aligned alternative to American AI governance.

Every governance action taken against Anthropic’s models has produced a response that reduces American AI’s structural advantage. Export controls → EU sovereignty push → domestic AI investment. Model gating → MoA routing around the gate. Company restriction → foreign recruitment.

This is the governance paradox: the tools available to control AI deployment are the same tools that accelerate AI’s global redistribution.

The answer isn’t “don’t regulate.” The answer — which OpenAI is partially implementing with Daybreak and Jalapeño — is build the governance infrastructure before the launch, not after. Pre-coordinated government review. Allied nation partnerships. Identity verification already in place. Then the “voluntary” review actually works, because there’s no adversarial relationship to trigger escalation.

Anthropic called for a global pause (June 5). Got a forced pause (June 12). Is negotiating its way out of that pause with identity verification and D.C. meetings. And watching Austria recruit it to Europe while the open-source community builds around the frontier it helped define.

The physical framework exists. The legal protections need to be strengthened.

That Digg comment, about how Azure Foundry and Bedrock already enable geographically isolated AI deployments, captures the real opportunity. The governance problem isn’t unsolvable. It’s just that the solutions require infrastructure investment before launch, relationship building before crisis, and governance architecture before capability crisis forces improvised export controls.

Day 160: Limited Mythos restoration happening quietly. Austria’s letter public. Hermes MoA beating the gated frontier at 80× the cost. China still accessing Claude through VPNs. OpenAI building the chip that will make inference 50% cheaper and harder to distill.

The era of self-regulated AI deployment ended June 25, when the government asked OpenAI to stagger GPT-5.6 and OpenAI said yes. What replaces it is still being negotiated, country by country, company by company, directive by directive.

That negotiation is what the next 160 days will be about.

Keywords for June 29, 2026:

1. Recruitment
2. Routing
3. Paradox
4. MoA
5. Redistribution