At 5:21 PM Eastern on Friday June 12, 2026, the United States government told Anthropic to stop selling its two most capable AI models to anyone who is not an American citizen. The order was not theoretical. The letter, signed by Commerce Secretary Howard Lutnick and addressed to Anthropic chief executive Dario Amodei, was delivered the same day SpaceX began trading on Nasdaq and three days after Fable 5 became the first publicly available Mythos-class model in the world. Anthropic could not, in real time, verify the citizenship of every person using its API. The company chose compliance. Fable 5 and Mythos 5 are now offline for every customer on Earth.
Anthropic’s public statement on the directive, posted to X at 12:50 AM Saturday New Zealand time, has 78.7 million views, 83,290 likes, 24,705 reposts. The post is the most-read AI announcement of the year by an order of magnitude. The engagement is not because of the product. It is because of what the directive means. The United States has, for the first time in history, placed an online API — a text box on a web page — under the same legal regime that governs the export of fighter-jet components, encryption software, and weapons-grade semiconductors. Every other AI model in America is now operating in the legal shadow of that decision.
🔍 THE BOTTOM LINE
The US government, citing national security but providing no specific technical detail, ordered Anthropic on Friday evening to suspend all foreign access to Fable 5 and Mythos 5. Anthropic, unable to verify citizenship in real time, suspended both models for every customer in the world. The company publicly disagrees with the directive. The government’s apparent concern is a jailbreak method Anthropic describes as “a small number of known, minor vulnerabilities” that other public models also have. The directive does not name the jailbreak, does not name the threat actor, does not describe the national security risk in any technical detail, and applies to API access — not chip exports, not model weights, not hardware. An input box on a web page is now an export-controlled capability. That is the precedent.
The Timeline
The week that ended with Fable 5 and Mythos 5 going dark was the most consequential week in the public life of frontier AI so far. To understand the directive, you have to see it in context.
Tuesday June 9, 2026. Anthropic released Claude Fable 5, the first Mythos-class model made publicly available. It scored 80.3% on SWE-Bench Pro. Stripe used it to migrate a 50-million-line codebase in one day. The launch included the related Claude Mythos 5, restricted to vetted Project Glasswing partners, focused on cybersecurity and biology research. By June 12, the Glasswing program had expanded to about 150 organizations in 15 countries.
Wednesday June 10, 2026. Google DeepMind posted the 60-page paper “From AGI to ASI” to arXiv, with Shane Legg, Marcus Hutter, and Iason Gabriel among the authors. The paper argues that AGI is unlikely to be a single event and that the transition to ASI will be a cascade. The same week, the world’s most senior AI research team and the world’s most senior US AI lab were each independently putting their cards on the table.
Thursday June 11, 2026. Reports circulated that Demis Hassabis would propose the “Einstein Test” — train an AI on pre-1905 physics, see if it can derive relativity — placing AGI “years, not months” away. The credibility benchmark became the talking point of the day.
Friday June 12, 2026, 9:30 AM ET. SpaceX began trading on Nasdaq under the ticker SPCX at a $135 opening price and a $1.77 trillion valuation. The largest IPO in history. Anthropic’s IPO is now expected in October; OpenAI’s in early 2027.
Friday June 12, 2026, 5:21 PM ET. Commerce Secretary Howard Lutnick’s letter arrived at Anthropic’s headquarters. It invoked national security authorities. It required Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national, “whether inside or outside the United States, including foreign national Anthropic employees.” Per Axios, the letter placed the two models under export control restrictions covering overseas access, re-export, and domestic transfer to foreign persons within the United States. The letter did not provide specific details of the national security concern. Per Anthropic’s own statement, the company’s understanding is that the government believes it has become aware of a method of jailbreaking Fable 5. Anthropic says the jailbreak it was shown involved only a small number of known, minor vulnerabilities that other public models can also achieve.
Friday June 12, 2026, evening. Anthropic made the only defensible call. The company could not reliably verify, in real time, which API users were foreign nationals and which were not. To ensure full compliance, it disabled both models for every customer in the world. Access to all other Claude models — Opus 4.8, Sonnet 4.6, Haiku 4.5 — is unaffected.
Saturday June 13, 2026, 12:50 AM NZT. Anthropic posted the public statement on X. 78.7 million views in the first 30 hours.
The Jailbreak Question
The most consequential sentence in Anthropic’s statement is the one about the underlying reason. The government, the company says, “has become aware of a method of jailbreaking” Fable 5. Anthropic considers the demo it saw to “involve only a small number of known, minor vulnerabilities that other public models can also achieve.” The directive was therefore triggered by a vulnerability that, in Anthropic’s view, is not unique to Fable 5 and not materially worse than what already exists in the public AI stack.
If that framing is accurate, the directive is not about the jailbreak. It is about the precedent. The US government is signalling that, in its view, certain categories of AI capability warrant export control regardless of whether the same capability is available elsewhere. Fable 5 is the first Mythos-class model released into the public market. It is the natural place to draw the line.
There are two ways to read this. The first is that the government is acting on a genuine but undisclosed intelligence concern — a nation-state or criminal actor has demonstrated a Fable 5 jailbreak chain that materially lowers the cost of cyber operations, and the directive is a precaution. The second is that the directive is a political move: an administration that has staked its AI policy on American leadership is drawing a hard line on the most capable US-built model, in the week of the largest IPO in history, on the eve of the G7 summit where sovereign AI is on the official agenda. Both readings can be true. The directive, like most national security directives, contains no technical detail and no due process for the affected company. Anthropic is contesting it publicly, which is the only lever it has.
The API as a National Security Capability
The other precedent — and this is the one the legal scholars will spend the next decade arguing about — is the expansion of export controls from physical artefacts to online services. Export controls have, historically, covered things. Chips. Encryption software. Machine tools. Satellite components. Model weights, since 2023, are covered under the Biden-era AI diffusion rule. API access — the act of sending a prompt to a model and getting a response back — has never been.
The Lutnick letter changes that. Under the directive, the act of typing a question into a web form and receiving an answer is now treated as a controlled capability transfer. If you are a foreign national, you cannot lawfully receive the output of Fable 5. If you are a US person, you can. If you are a foreign national sitting in a San Francisco coffee shop using your US employer’s laptop, you cannot. If you are a foreign national Anthropic employee, you cannot use your own company’s product. The instruction reaches inside the company itself.
The implications cascade. Every US AI lab now faces a question it has not previously had to ask: at what capability level does our model become a controlled technology, and what is the compliance path? The answer, for now, appears to be: at the Mythos-class level, and there is no clear compliance path. Anthropic chose to suspend the models entirely. Other labs will face the same choice the first time they ship a model that crosses the line.
What Anthropic Is Saying, And What It Is Not
The official Anthropic statement is careful. The company says it “apologises for this disruption to our customers.” It says it “believes this is a misunderstanding and is working to restore access as soon as possible.” It does not say the directive is wrong. It does not say the government is acting in bad faith. It does not say the jailbreak is fictional. It says only that the jailbreak is minor and the response is disproportionate.
That restraint is deliberate. Anthropic has a multi-billion-dollar relationship with the US federal government. In July 2025, the Department of War awarded Anthropic a two-year, $200 million contract to prototype frontier AI capabilities for national security. The company’s public position on AI policy has been, since at least October 2025, that AI governance should be “a matter of policy over politics” and that the company is “aligned with the Trump administration on key areas of AI policy.” Anthropic’s CEO Dario Amodei wrote in October that “managing the societal impacts of AI should be a matter of policy over politics.” The Friday directive is a test of that framing. The company’s response — comply, apologise, disagree quietly — is consistent with the position Amodei has been building for the last eight months.
The harder question is what Anthropic does next. If the directive stands, Fable 5 and Mythos 5 are functionally dead products for any non-US customer, and the company has to decide whether to build a parallel non-US version, a US-only version, or accept the constraint and ship a different model tier globally. None of those options is attractive. The first splits the model lineage. The second concedes the non-US market. The third demotes Mythos 5 from flagship to US-domestic.
What It Means for New Zealand
The first-order impact on New Zealand is small. Anthropic’s Fable 5 and Mythos 5 are restricted products. Their loss does not change the day-to-day experience of Claude users in Auckland or Wellington. The interesting effects are at the structural level.
New Zealand is the end of every global supply chain. When the US government puts an AI API under export control, the rule applies in New Zealand the moment the directive is signed. The directive does not need a New Zealand Parliament debate, a Foreign Affairs briefing, or a Privacy Commissioner ruling. The control is upstream. The country inherits the consequence. This is the same pattern as the HBM memory shortage hitting the Mac Studio supply: the policy is set in Washington or Beijing, and New Zealand consumers experience the outcome. The cascade framing from the DeepMind paper is not abstract. It is the operating reality for a small economy.
The second-order impact is more interesting. The directive establishes, for the first time, a legal precedent that AI API access is a national security capability subject to export control. The G7 summit in France starts Monday. The sovereign AI framework Macron is pushing assumes that AI infrastructure, models, and standards are under national control. The US directive is, in effect, the most concrete sovereign AI action any G7 nation has taken to date. Other G7 members will be asked, in the next 48 hours, whether they will follow. The answer to that question will determine whether the post-summit framework is multilateral (every G7 country has its own export controls on its own frontier models) or unilateral (the US controls its models, everyone else has to negotiate access).
For New Zealand, the multilateral outcome is better. A regime in which every G7 country has comparable restrictions means New Zealand can negotiate a single set of terms with a single set of counterparts. A unilateral US-only regime means New Zealand’s access to frontier AI depends on the terms the US chooses to offer, on whatever schedule the US chooses to update them, and the country has no seat at the table. The Singapore-Australia bilateral on AI data centres is the template for what a New Zealand-bilateral might look like. The G7 outcome on Monday will tell us whether that template is the right starting point or whether a harder conversation is coming.
The third-order impact is the local-AI story. If US AI APIs are now subject to discretionary foreign-access controls, the AMD $2,000 mini PC that runs a 235B model on-device stops being a niche option. It is the only AI stack that the US government cannot cut off with a letter. Open-weight models, running on local hardware, in a household or a school or a small business, are export-control-proof by construction. The directive strengthens every argument for local AI infrastructure investment — and weakens every argument for treating the cloud APIs as a free, stable, dependable foundation for serious work.
⚠️ THE OTHER SIDE
Three honest caveats. First, the national security concern is undisclosed. The government may have intelligence the public does not. A jailbreak chain that, in Anthropic’s view, is minor, may, in an intelligence analyst’s view, be the specific capability an adversary needed. We do not know. The directive gives Anthropic no opportunity to contest the technical basis, and gives the public no way to evaluate the proportionality of the response. Second, Anthropic has an interest in the directive being reversed. The company is in the middle of an IPO process and is filing to go public at a trillion-dollar valuation. A model that is restricted from sale to half the world’s developers is worth less than a model that is not. Anthropic’s public disagreement is consistent with its fiduciary interest, not just with the technical reality. Third, the precedent is not necessarily bad. Export controls on chips arguably delayed China’s domestic AI buildout by 18–24 months. Export controls on AI APIs may do the same. The question is whether the policy is targeted, time-limited, and reviewable, or whether it becomes a permanent feature of the AI market. The Lutnick letter gives no signal on that question. The G7 summit is the next data point.
❓ FREQUENTLY ASKED QUESTIONS
What happened to Fable 5 and Mythos 5? Both models are offline for every customer worldwide, including US customers. Anthropic disabled them on Friday June 12 evening, New Zealand time, in response to an export control directive from the US Commerce Department. Access to all other Claude models (Opus 4.8, Sonnet 4.6, Haiku 4.5) is unaffected.
Why did the US government order the suspension? The Commerce Department cited national security authorities. Per Anthropic’s public statement, the government’s stated concern is a method of jailbreaking Fable 5. Anthropic says the jailbreak it was shown involves only minor vulnerabilities other public models also have. The directive contains no specific technical detail.
Who signed the directive? Commerce Secretary Howard Lutnick. The letter was addressed to Anthropic CEO Dario Amodei and delivered at 5:21 PM Eastern on Friday June 12, 2026.
Does the directive apply to foreign nationals inside the United States? Yes. The directive covers all foreign nationals “whether inside or outside the United States, including foreign national Anthropic employees.” A non-US person using an Anthropic API in San Francisco, or working at Anthropic’s San Francisco headquarters, is covered.
Are OpenAI, Google DeepMind, or xAI models affected? No. The directive is specific to Anthropic’s Fable 5 and Mythos 5. OpenAI’s GPT-5.5, Google’s Gemini 3, and xAI’s Grok 4 are not subject to the directive and are operating normally.
Is this the first time an AI API has been put under export control? Yes. The US has previously put AI chips, model weights (under the 2023 AI diffusion rule), and certain dual-use AI capabilities under export control. This is the first time an online API — the act of sending a prompt and receiving a response — has been treated as a controlled technology.
What happens to Anthropic’s IPO? The IPO timeline is unaffected in the short term. The S-1 was filed on June 1. The pricing and listing date are set independently. A long-term suspension of Fable 5 would reduce the company’s revenue trajectory and the multiple the public market is willing to pay. A short-term suspension is recoverable. The directive’s duration is the key unknown.
How does this connect to the G7 summit? The G7 summit in Évian, France, runs June 15–17. Sovereign AI is on the official agenda. The US directive is the most concrete sovereign AI action any G7 nation has taken to date. Other G7 members will be asked, in person, whether they will follow. The answer to that question will determine the post-summit framework.
What should a New Zealand AI user do? Continue using Claude Opus 4.8, Sonnet 4.6, or Haiku 4.5 as normal. The directive does not affect those models. If you were specifically using Fable 5 or Mythos 5, those models are offline. For long-term planning, the local-AI hardware option — a $2,000 AMD mini PC running open-weight models on-device — is now structurally more attractive than it was a week ago. Open-weight models cannot be cut off by a foreign export-control directive.
Will the directive be reversed? Unknown. Anthropic is working to restore access. The Commerce Department has not signalled whether the directive is permanent, time-limited, or contingent on technical conditions being met. The next data point is the G7 summit, where the question of multilateral export controls will be on the table for the first time.
