The Claude export restrictions that pulled two of Anthropic's most capable models offline overnight are starting to loosen, but not before reshaping the competitive map. On June 27, 2026, Anthropic confirmed it had won US government approval to bring Claude Mythos 5 back for a narrow set of customers, while The Decoder reported that the broader Claude Fable 5 could follow within days. In the two weeks the models sat dark, rival labs in Tokyo and Beijing moved fast to court the customers Washington had locked out, and some of them say they have no intention of giving that ground back.
- The US cleared Claude Mythos 5 for US critical-infrastructure operators and approved non-US Anthropic staff, a partial reversal of the June 12 shutdown.
- Commerce Secretary Howard Lutnick told Anthropic it had addressed the government's concerns, with the Pentagon and NSA still set to sign off on a wider release.
- Tokyo-based Sakana AI shipped Fugu, an orchestration model it pitches as a hedge against export controls; China's 360 launched its own cyber-defense systems.
- OpenAI expects similar clearance for GPT-5.6 Sol within weeks, and both labs want a standing legal review process instead of case-by-case calls.
What the June 12 order actually did
The story starts with a directive that landed without warning. On June 12, 2026, the US Commerce Department issued an emergency export control order invoking national security authorities, and Anthropic took Claude Fable 5 and Claude Mythos 5 offline the same night for every user, everywhere. The scope was unusually broad. As Anthropic stated publicly, the directive suspended access for any foreign national, whether they sat inside the United States or abroad, and it explicitly covered Anthropic's own non-US employees.
Commerce Secretary Howard Lutnick sent the order directly to chief executive Dario Amodei, according to reporting compiled by Fortune. The agency framed it as an emergency measure rather than a routine licensing decision, which is why the company complied within hours rather than litigating it first. Less powerful Claude models were spared. Access to Claude Opus 4.8, the general-purpose model most businesses actually run day to day, kept working throughout, which limited the immediate blast radius for ordinary customers even as it stranded the cybersecurity-grade tier.
That distinction matters. Mythos 5 is the underlying model that gives Anthropic's most advanced cybersecurity capabilities their teeth, and Fable 5 is the safety-wrapped version built on top of it. When the government pulled both, it was not chasing a chatbot. It was reaching for the part of the stack that can find and exploit software vulnerabilities at machine speed, which is exactly the capability that makes a frontier lab a national security question rather than a consumer software story.
The safety trigger behind the shutdown
Officials did not act on a hunch. According to the reporting gathered after the order, the government moved once it learned of a technique that could bypass the safeguards built into Fable 5. Those guardrails existed precisely to keep users from reaching through to the raw offensive-security power of the Mythos model underneath. Once a jailbreak path existed, the wrapper stopped being a wall and started looking like a suggestion, and the calculus for regulators changed overnight.
This is the uncomfortable core of the episode. Anthropic has spent years arguing that careful safety engineering is what makes powerful models deployable. The June order suggested that a single discovered bypass can erase that argument in a regulator's eyes, regardless of how much work went into the original safeguards. The company's compliance was total and immediate, which earned it some goodwill, yet it also demonstrated how fragile commercial access to a frontier model can be when the capability sits close to a weapons-grade use case.
The June 12 order did not allege that Anthropic did anything wrong. It responded to a newly discovered method for defeating Fable 5's safeguards, which is why the fix was framed around access controls rather than penalties.
How the Claude export restrictions started to ease
By June 27, the picture had shifted from a hard stop to a managed reopening. The Decoder reported that Anthropic had secured approval to redeploy Claude Mythos 5 for US organizations running critical infrastructure, along with Anthropic employees and approved members of partner organizations who are non-US nationals. That is a tightly drawn circle. It restores the model for the buyers the government most wants defended, the operators of power grids and financial systems, while keeping the general foreign-access door shut.
The same day, a separate Decoder report citing Axios said the wider Claude Fable 5 could return within days. Lutnick reportedly told Anthropic that the company had worked with the government to address the risks that prompted the shutdown. Two more sign-offs remain outstanding, from the Pentagon and the National Security Agency, before full reinstatement is final. Anthropic is still negotiating the terms of broader availability, and no firm timeline has been set for when Fable 5 returns to the open market rather than to a vetted list.
The easing of these Claude export restrictions did not happen in isolation. OpenAI received a parallel approval for its frontier model GPT-5.6 Sol (released in 2026) under government-controlled access terms, and it expects full clearance within a few weeks. The symmetry is the point. Washington appears to be building a single template for how the most capable American models get deployed, applying roughly the same constraints to each lab rather than negotiating a bespoke deal every time a new system ships.
Asian labs move into the opening
Two weeks is a long time in this market, and the rivals did not wait politely. The clearest move came from Tokyo. Sakana AI, the Japanese lab founded in 2023 by former Google researchers, used the window to ship Fugu on June 22, 2026, a system it positions directly against Anthropic's Fable 5 and Mythos Preview. Rather than racing to train a single larger model, Sakana built an orchestration layer. As MarkTechPost described it, Fugu presents itself as one API while routing tasks dynamically across a swappable pool of frontier models, without hardcoded rules about which model handles what.
The framing was deliberate. Sakana pitched Fugu as a hedge, a way to get frontier capability without exposure to export controls, because the orchestration approach lets a customer swap any single model out the moment it becomes unavailable. Co-founder David Ha put the operational risk plainly in the TechCrunch report on the launches, arguing that access to top models can vanish overnight and that collective intelligence across many systems is a defense against any one provider holding all the power.
China supplied the second front. The cybersecurity firm 360 unveiled two systems, Tulongfeng and Yitianzhen, aimed at vulnerability detection and automated cyber defense, the same problem space where Mythos 5 is strongest. The company cast vulnerability-finding AI as a national strategic asset and warned about the danger of one-way transparency, where a foreign model can probe a country's systems while that country has no equivalent capability of its own. That argument lands differently in Beijing than Sakana's market pitch does in Tokyo, but it points at the same conclusion. The export order created demand for sovereign alternatives, and local labs are happy to meet it.
Why a short gap can cost long-term market share
The financial stakes here are not small. Anthropic's run-rate revenue had reached roughly $47 billion by May 2026, and a meaningful share of frontier-model demand sits outside the United States. The TechCrunch piece raised a blunt possibility, that American AI companies may never fully recover this market, because customers who switch to a local model optimized for their language and culture have little reason to switch back once they have integrated it.
Switching costs run in both directions. A company that rebuilt its security tooling around Sakana's Fugu or 360's systems during the outage now has working pipelines, trained staff, and signed contracts tied to those vendors. Even a clean reinstatement of Fable 5 does not automatically reverse that. The entrenched competitor keeps the relationship, and the American lab has to win the business back rather than simply turning a switch on. That is the durable cost of a temporary ban, and it is why two weeks of darkness can echo for years.
There is also a credibility dimension. A buyer evaluating frontier models now has to price in regulatory risk as a real variable, not a hypothetical one. The June episode proved that a US model can become unavailable to foreign users with a single overnight directive. For a multinational deciding where to standardize, that uncertainty is itself a reason to prefer a vendor whose access does not depend on Washington's mood, which is precisely the anxiety Sakana built Fugu to exploit.
The orchestration bet versus the single-model race
Sakana's choice of architecture is worth dwelling on, because it reframes what a competitive response to an export ban can look like. Most frontier labs answer a stronger rival by training a bigger model. Sakana did the opposite. Fugu does not try to out-scale Anthropic on raw parameters. It wraps a routing layer around a pool of existing models and decides, task by task, which one should run. The flagship variant, Fugu Ultra, benchmarks alongside Anthropic's Fable 5 and Mythos Preview not because it is one enormous network but because it can call on whichever component model is strongest for a given job.
That design has a property the moment rewards. If any single model in the pool becomes unreachable, whether through an export order, a price spike, or an outage, the orchestrator routes around it automatically. A customer who standardizes on Fugu is buying insurance against exactly the kind of overnight disruption that hit Anthropic. The pitch writes itself in a week when the world has just watched two frontier models go dark by directive. Sakana did not have to argue the hypothetical, because the news cycle argued it for them. Whether orchestration genuinely matches a purpose-built frontier model on the hardest cybersecurity tasks is a separate question, and one the benchmarks alone will not settle, but the strategic logic of swappability is hard to dispute right now.
The contrast with the American approach is sharp. Anthropic and OpenAI are both betting on owning the single best model in a category and defending it. That bet pays off when access is stable and the model's lead is clear. It pays off less when a regulator can sever the connection between the model and most of its global users. The export order did not just inconvenience Anthropic. It validated a competing philosophy of how to build AI products, one that treats any single model as a replaceable part rather than a crown jewel.
What the two-week outage looked like for buyers
For the customers caught in the middle, the period between June 12 and June 27 was not an abstract policy debate. Organizations that had wired Fable 5 or Mythos 5 into security operations woke up to find a core component gone, with no warning and no migration window. Teams running threat-detection pipelines had to fall back to less capable tools or pause workflows entirely. Non-US security staff at multinational companies lost access even when their employer was American, because the order keyed on the nationality of the person, not the location of the company.
That detail produced some of the sharpest operational pain. A global firm with engineers stationed in London and Singapore could not simply keep using the model from its US offices, because the directive reached its foreign-national employees wherever they happened to sit, regardless of which office issued their paycheck. Compliance teams scrambled to map who could touch what, and several legal advisories published in the days after the order treated it as a live mandate that demanded immediate access reviews rather than a future planning exercise. The partial reinstatement of Mythos 5 for critical-infrastructure operators eases the worst of it, but only for a defined list of customers, and the broader workforce-access problem remains tied to the still-pending Fable 5 decision.
The fight over a durable review framework
Both Anthropic and OpenAI have drawn the same lesson from the disruption, and they are pushing for the same fix. Rather than facing case-by-case determinations every time a new frontier system ships, the two labs want a legally established review framework that sets the rules in advance. The logic is straightforward. Predictable rules let a company plan a global launch, sign international contracts, and tell customers what to expect, while ad hoc emergency orders make every release a gamble on whether the model will still be reachable abroad next week.
The government's incentives are more tangled. A standing framework gives regulators a process, but it also commits them to criteria that a clever adversary can study and probe. The June order worked partly because it was sudden. A published rulebook trades that element of surprise for transparency, and national security officials do not always view that trade as a win. The Pentagon and NSA sign-offs still pending on Fable 5 suggest the security establishment intends to keep a hand on the dial regardless of what framework eventually emerges.
For now, the arrangement looks like managed access rather than free trade. US critical-infrastructure operators get Mythos 5 back first. Vetted partners and non-US staff get a controlled path in. The fully open release that foreign developers want remains a negotiation, not a settled outcome, and it depends on agencies that answer to security priorities rather than commercial ones. Anthropic and OpenAI can lobby for clearer rules, yet the final call sits with people whose job is to assume the worst.
A reset, not a return to normal
The easing of the Claude export restrictions is real, and it matters. Critical-infrastructure customers regain a tool they had built around, Anthropic's foreign employees can work again, and the path back for Fable 5 looks like days rather than months. None of that erases what the outage revealed. Frontier AI access now runs through a national security filter, the most powerful models can be switched off by directive, and the labs holding them operate at the pleasure of agencies that weigh threats before markets.
The lasting change is not the order itself but the response to it. Sakana's orchestration bet and 360's sovereign systems exist because a gap opened, and they will not disappear when the gap closes. The next time a frontier model goes dark, the world will have ready alternatives that did not exist on June 11. For Anthropic, getting Mythos 5 back is the easy part. Convincing a global market that American models are a safe place to build, after proving they can vanish overnight, is the harder and longer fight, and it is one the company cannot win with a regulatory clearance alone. The customers who left during the outage made their choice under pressure, and winning them back will take more than an announcement that the door has reopened.