A two-and-a-half-hour lunch in Évian-les-Bains on Wednesday was supposed to be about how AI drives economic growth and keeps societies resilient for young people. Officially, it still is. Unofficially, the Politico EU report on the G7 AI summit lunch makes clear the room’s real subject is the one nobody put on the agenda: the US directive that blocked non-US citizens from accessing Anthropic’s Fable 5 and Mythos 5 models, forcing Anthropic to cut global access to its most tightly controlled cyber-capable systems last Friday.

🔍 THE BOTTOM LINE

Europe is publicly miffed at the Anthropic cutoff and privately treading very carefully. The G7 lunch is the first face-to-face between EU leaders, Donald Trump, and Anthropic CEO Dario Amodei since Washington pulled the lever — and the choreography of “engagement” is doing real diplomatic work to keep the relationship intact. But underneath the bonhomie, the structural fact is now visible to every AI minister in Europe: frontier model access is US-leveraged, and the export-control regime that produced the Anthropic block can be aimed at anyone.

What Changed

The Anthropic block itself was a Friday event. What changed this week is that European leaders had to decide how to handle it in person. Three signals from the summit, per Politico’s reporting:

• EU tech chief Henna Virkkunen told the European Parliament on Tuesday that “contingency measures taken in this light should not be discriminatory against partners” — the closest any senior EU figure has come to a public objection. No name attached. No escalation.
• Commission President Ursula von der Leyen’s bilateral with Trump on Tuesday did not mention the Anthropic issue at all, according to the read-out. The Commission’s deputy chief spokesperson Arianna Podestà refused to preview Wednesday’s lunch: “I will not be prejudging now what will be raised or not.”
• Anthropic is sending Amodei to a follow-up meeting with Commission and EU cyber authorities in San Francisco on Thursday — the Anthropic statement, quoted by Politico, frames the visit as “ongoing engagement with the EU, allied democracies and important international institutions on frontier AI’s implications for cybersecurity.”

The four CEOs at the lunch — Amodei, OpenAI’s Sam Altman, DeepMind’s Demis Hassabis, and Mistral’s Arthur Mensch — were officially there to discuss growth and resilience. Anthropic confirmed Amodei’s attendance but declined further comment on the G7 discussions. One industry representative called the Anthropic-US row “the elephant in the room.”

Context

The Anthropic cutoff landed without formal US government communication — only the Friday statement Anthropic itself issued. That detail matters: it means there was no inter-government notification, no negotiating window, no advance warning to allied capitals. Brussels is treating the directive as a fait accompli to manage around, not a policy to contest head-on.

The EU’s AI Act already imposes risk-testing and evaluation obligations on frontier model providers operating in the bloc. The Anthropic block exposes a structural asymmetry: the EU can regulate the behaviour of frontier AI within its borders, but it cannot secure access to those models when Washington decides they shouldn’t leave US hands. The two policy levers point in opposite directions, and Europe only controls one of them.

Italy’s Giorgia Meloni is using the lunch to push a line she has been developing since Rome chaired the G7 in 2024: AI “responsibility” and the definition of truth in artificial intelligence, including watermarking AI-generated content. An Italian diplomat said Meloni raised the same theme with Japanese Prime Minister Sanae Takaichi in Rome on Monday.

The Other Side

Not every European tech executive at the lunch is publicly scolding Washington. Domyn CEO Uljan Sharka, one of the few European founders in the room, defended the US move to Politico: “I don’t blame the US administration for doing that. They were pushed and forced to take action” — a reference to Anthropic’s branding of its models as elite cyber-capable. Sharka added: “This narrative of us versus them is completely wrong, and I hope that during the G7 this is going to be addressed.” He argued for “NATO-style” cooperation on AI alongside the existing transatlantic defence partnership.

UK AI Minister Kanishka Narayan took a sharper line: “The main lesson,” he said, “is that as we debate the future of national security and technological sovereignty, access to AI capabilities is crucial.” The UK government has been in touch with both Washington and Anthropic “to understand the situation.” Narayan did not name any specific policy response.

NZ Angle

New Zealand sits exactly where Europe was ten days ago: a small allied democracy running almost entirely on US frontier AI access, with no domestic frontier-model capability, no seat at the G7 table, and no advance warning infrastructure if Washington decides to extend the export-control regime. Te Herenga Waka—Victoria University of Wellington has one of the country’s stronger AI policy research clusters, the NZ AI Forum has published on sovereign capability, and the government’s emerging AI strategy still leans heavily on “trusted partner” arrangements with US labs. None of that provides leverage when the lever is pulled.

The lesson from the G7 choreography is that “engagement” buys time, not access. Europe had two years of EU AI Act negotiations, a €50 billion-plus compute investment programme in EUROMESH, and a permanent seat at every major AI table — and it still got locked out of Fable 5 with no formal notice. New Zealand has none of those buffers. A directive similar to the Anthropic one, aimed at any New Zealand-relevant provider, would not even produce a press cycle.

The Bigger Picture

The G7 lunch is unlikely to produce a public statement on the Anthropic block. European officials have spent the week lowering expectations to “recreating a circle of trust” — the phrase one European diplomat used with Politico. That is the right diplomatic register for an ally that does not want to publicly fight with the US while dependent on US AI. It is the wrong register for what is actually happening.

What is actually happening is that the United States has discovered it can use the export-control regime as a tool against allied populations, and the Anthropic block is the proof of concept. The first use of any new coercive mechanism is calibrated, deniable, and aimed at an actor — Anthropic — that is unlikely to escalate. The next use may be less polite. The G7’s “engagement” framing is a polite response to a rude mechanism, and the politeness will not be what stops the next use.

❓ FAQ

Did any G7 leader publicly confront Trump or Amodei about the Fable 5 block at Wednesday’s lunch? No. Virkkunen’s Tuesday remarks to the European Parliament were the closest Europe came to a public objection, and she named no country. The von der Leyen–Trump bilateral on Tuesday did not mention the issue at all.

What was Anthropic’s role in the cutoff — was the company forced, or did it choose to comply? Anthropic confirmed on Friday that it had cut off global access to Fable 5 and Mythos 5 after the US directive. Its Wednesday statement on the SF follow-up framed the cuts as compliance with the export-control regime and emphasised “ongoing engagement” with EU institutions.

Why is this a sovereignty issue for Europe, not just an Anthropic commercial problem? Because the cutoff shows that frontier AI access is now a US-leveraged policy variable. The EU can regulate how AI behaves inside its borders, but cannot secure access to frontier models from outside them. That asymmetry is a strategic problem independent of Anthropic specifically.

Did Arthur Mensch (Mistral) comment on the Anthropic block at the lunch? Politico did not quote Mensch on the Anthropic issue. Mensch was present as one of the four CEO attendees and his attendance matters for a different reason: Mistral is the largest European frontier-model developer, and his presence in the room signals Paris’s interest in being treated as a sovereign AI counterweight.

What happens at the Thursday San Francisco meeting between Anthropic and EU cyber authorities? The official framing, per Anthropic’s statement to Politico, is cybersecurity and frontier-AI implications. The actual agenda is undisclosed. The fact that the meeting is happening at all signals both sides want a working channel that survives the export-control row.

Sources

• Politico EU — Europe wary of stoking Anthropic row at G7 (Clea Caulcutt, Giorgio Leali, Joseph Bambridge)
• Wired — Anthropic Is Still at Odds with the White House over Claude Fable 5
• Reuters — OpenAI, Anthropic, Google DeepMind policy coverage
• Bloomberg — G7 AI summit reporting
• The Guardian — EU AI policy and export controls
• Singularity.Kiwi — EUROMESH and European sovereign AI compute