Dario Amodei gets Fable 5 back online after Anthropic's export-control scare

Why it matters

Anthropic got Fable 5 back online, but the episode turns model access into a board-level platform risk: frontier capability now depends on safety evidence, government process, and cloud distribution as much as benchmarks.

Dario Amodei and Daniela Amodei's Anthropic is restoring global access to Claude Fable 5 on Wednesday, July 1, after the U.S. government lifted export controls that had forced the company to pull its two newest models from users worldwide.

The return, detailed in a June 30 Anthropic post, ends the immediate commercial damage from a 19-day shutdown of Fable 5, Anthropic's safeguarded general-use frontier model. It does not return the company to the pre-June 12 status quo. Anthropic is bringing Fable 5 back with a new cyber-safety classifier, a more formal government testing pipeline, and a proposed industry framework for scoring model jailbreaks with Amazon, Microsoft, Google, and other Project Glasswing partners.

That is the Amodeis' founding bargain under pressure. Anthropic was built by former OpenAI researchers around the claim that frontier models could be made more reliable, steerable, and safe enough to deploy. Dario, a Princeton-trained physicist who later led research at OpenAI and helped drive GPT-2 and GPT-3 work, has made controlled release part of the company's identity. Daniela, Anthropic's president and a former Stripe employee before her OpenAI policy role, emphasized in a 2023 Stripe interview the goal of making AI systems safer from day one.

Fable 5 is the clearest test yet of whether that thesis can survive contact with export-control law, cloud distribution, and customers who expect frontier AI access to be durable.

What returns on July 1

Anthropic says Fable 5 will be available globally on Claude Platform, Claude.ai, Claude Code, and Claude Cowork starting Wednesday, July 1. For Pro, Max, Team, and select Enterprise plans, Anthropic says Fable 5 will count for up to 50% of weekly usage limits through July 7. After that, access moves to usage credits, the paid overage system that bills additional Claude use at standard API rates and carries a $2,000 daily redemption limit.

Cloud distribution is still not fully back. Anthropic says it will re-enable access on AWS, Google Cloud, and Microsoft Foundry "as quickly as possible," but it did not give a date. That gap matters because Anthropic's enterprise strategy depends on meeting customers where their compute and compliance stacks already live. In May, the company said Claude was the first frontier model available on Amazon Web Services, Google Cloud, and Microsoft Azure, and that AWS remained its primary cloud provider and training partner.

Anthropic also says Mythos 5 access has been restored for a set of U.S. organizations after government approval on June 26. Mythos 5 is not the general product. It shares the same underlying model as Fable 5, but Anthropic says it has fewer safeguards and is meant for vetted defensive cybersecurity users in Project Glasswing.

Why the controls landed

Anthropic released Claude Fable 5 and Claude Mythos 5 on June 9. Three days later, on Friday, June 12, the U.S. government applied export controls to both models. Anthropic says the directive required it to restrict access to foreign nationals both inside and outside the United States. Because the order took effect immediately and Anthropic says it lacked a reliable real-time nationality verification system, it suspended both models for all users.

Anthropic says the government action followed a report from Amazon researchers who found a way to bypass Fable 5's safeguards so the model identified software vulnerabilities. In one case, according to Anthropic, the model produced code showing how a vulnerability could be exploited.

The Amazon report itself was not published with Anthropic's announcement, and the legal instrument behind the June 12 control was not included in the post. Those omissions limit what can be independently assessed. The public evidence is Anthropic's account of the review, Howard Lutnick (@howardlutnick)'s post saying the controls were lifted, and outside legal analysis of the restriction.

Anthropic's core defense is that the reported behavior was not unique to Fable 5 or Mythos 5. The company says its testing found that less capable models, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the same vulnerabilities described in the Amazon report. Anthropic also says every model it tested could produce the same exploit demonstration as Fable 5, including older Claude models, GPT-5.4, GPT-5.5, and Kimi K2.7.

That claim, if borne out by broader testing, cuts to the policy problem: the government did not merely restrict a specific open release or a leaked model weight. It restricted remote access to a commercial AI service because of a jailbreak allegation around cyber capability, even though the underlying behavior may be common across modern coding models.

The fix: stronger blocking, more false positives

Anthropic says it trained an improved safety classifier with the government and Amazon that targets the behavior described in the Amazon report. When Fable 5 blocks a request, Anthropic says the request will instead be sent to Claude Opus 4.8.

The company says the new classifier blocks the specific Amazon-reported technique in more than 99% of cases. It also acknowledges the cost: more benign coding and debugging requests will be flagged.

That tradeoff is not incidental. Anthropic says Fable 5 launched with conservative cyber-safety safeguards, meaning it intentionally blocked some likely harmless requests to reduce the chance of harmful output. The June 30 changes widen that operational bargain. Developers get Fable 5 back, but some of their requests will be redirected to a less capable model when the classifier sees cyber risk.

For users, that can feel like random product degradation. For Anthropic, it is the price of keeping a model online after the government has shown it is willing to intervene.

Mythos remains a gated cyber weapon and shield

The harder question is Mythos 5. Anthropic says Mythos 5 can find and exploit software vulnerabilities more effectively than any other model and all but the most skilled human security experts. That is why the company has tried to put Mythos inside Project Glasswing, a controlled-access effort for cyberdefenders and critical infrastructure operators.

Glasswing had already become the center of Anthropic's cyber-safety argument before the Fable 5 fight. On June 2, Anthropic said it was expanding Project Glasswing to approximately 150 new organizations in more than 15 countries, after roughly 50 initial partners used Mythos Preview to scan codebases. The company said those partners had found more than 10,000 high- or critical-severity security flaws.

RuntimeWire reported on June 16 that security leaders viewed the Fable and Mythos restrictions as a risk to the same bug-finding work defenders need. Two days earlier, we noted how Anthropic's case for guarded access had become entwined with warnings from security researcher Nicholas Carlini, who had argued that the question was not simply release or ban, but who gets access and under what controls.

That is still the live issue. Anthropic says Mythos 5 is available again to a set of U.S. organizations, but broader domestic and international Glasswing access remains subject to government coordination.

The export-control precedent did not disappear

The controls were lifted, but the precedent remains. A CSIS analysis published June 16 said the Department of Commerce letter required an approved export license for foreign persons to access Fable 5 and Mythos 5, including foreign nationals inside the United States. CSIS also noted that the letter had not been made public and argued that Commerce's authority to control remote model access worldwide was uncertain.

That uncertainty is the market risk. If a frontier model can be pulled from global customers by a targeted government letter after a safety report from a cloud partner or rival lab, every enterprise buyer has to price access continuity into its AI stack. The obvious beneficiaries are smaller open-weight models, sovereign AI efforts, and non-U.S. providers that can sell jurisdictional certainty even when they cannot match the raw capability of Anthropic, OpenAI, or Google DeepMind.

Anthropic is trying to answer that risk with process. In the June 30 post, the company proposed a shared severity framework for AI jailbreaks.

The company is also committing to deeper government collaboration: pre-release access for designated government evaluators, faster information sharing on safeguards, dedicated resources for joint AI-security research, and work toward a common industry bar for frontier model providers.

That is not a retreat from commercialization. It is Anthropic's attempt to make regulation part of the release machinery before a future model gets stopped at the gate.

The money behind the patience

Anthropic can afford to take that route because it has raised enough capital to treat safety process as infrastructure, not overhead. On May 28, the company said it raised a $65 billion Series H at a $965 billion post-money valuation, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. The round also included $15 billion of previously committed hyperscaler investment, including $5 billion from Amazon.

The same announcement said Anthropic's run-rate revenue crossed $47 billion in May and that the funding would expand safety and interpretability research, compute, products, and partnerships. Those numbers are company-reported, but they explain why the Fable 5 disruption mattered. Anthropic is no longer a lab asking whether controlled deployment is possible. It is a near-trillion-dollar private company selling access to increasingly powerful models through consumer subscriptions, developer tools, and the three dominant clouds.

That scale changes the Amodeis' original safety argument. In 2021, when Anthropic announced its $124 million Series A, the company said it was building prototypes of reliable and steerable AI systems. In 2026, reliability means something larger: a customer has to know whether a model will be available tomorrow, whether a national-security review can interrupt a product line, and whether a false positive in a cyber classifier will route a task to a weaker model.

Fable 5's return is a win for Anthropic. It also marks a new operating reality for frontier AI founders. The product boundary is no longer set only by benchmark scores, pricing, latency, or enterprise demand. It is set by the founder's ability to convince cloud partners, security researchers, customers, and the U.S. government that the model can be powerful without becoming uncontrollable.