Anthropic nears US deal to restore Fable 5 and Mythos 5 access

Why it matters

The Anthropic talks could turn a one-off export-control shock into a de facto release process for every frontier AI model with advanced cyber capability.

Anthropic co-founder Tom Brown is helping steer negotiations with the Trump administration toward a deal that could lift US export restrictions on Anthropic's top two AI models, Fable 5 and Mythos 5, Bloomberg reported Friday, citing people familiar with the talks.

The talks center on whether Commerce Secretary Howard Lutnick and other administration officials can be satisfied that Anthropic has addressed security concerns around the systems. According to Bloomberg, Brown met with Lutnick and other senior officials in recent days, while Anthropic CEO Dario Amodei has taken a hands-off role in the negotiations. That personnel choice is the story inside the story: Anthropic's most important government fight is being handled not by its most publicly visible founder, but by another co-founder as the company tries to convert a confrontation into an access agreement.

The restriction was imposed on June 12, 2026, when Commerce directed Anthropic to suspend all access to Fable 5 and Mythos 5 by any foreign national, including foreign-national Anthropic employees. In a company statement that day, Anthropic said the order forced it to disable both models for all customers to ensure compliance, while leaving access to other Anthropic models unaffected.

Anthropic's public position was unusually direct. The company said the government had not provided specific details of its national security concern and that Anthropic understood the issue to be a method for bypassing, or jailbreaking, Fable 5. Anthropic said it reviewed a demonstration involving a small number of previously known, minor vulnerabilities and argued that other public models could find the same vulnerabilities without a bypass. It also said it had worked with the US government, the UK AI Security Institute, third-party groups and internal teams for thousands of hours of red-team testing before launch.

The central technical dispute has not changed: whether a narrow jailbreak risk is enough to justify blocking access to an entire frontier model. Anthropic says perfect jailbreak resistance is not currently possible for any provider, and warned that applying that standard across the industry would effectively halt new model deployments by frontier labs. Commerce has not publicly released the full technical basis for the order.

The timing matters for more than Anthropic's customers. Bloomberg reported that Anthropic recently filed confidentially for an initial public offering and that its latest valuation topped $900 billion. If the government can force a global shutdown of a flagship model two weeks after launch, public-market investors will have to price not only compute costs and model performance, but also the possibility that model access becomes a negotiable privilege rather than a product launch decision.

A voluntary framework is becoming an approval process

The Anthropic standoff is now setting the operating template for the rest of the frontier AI market. On June 2, President Donald Trump signed an executive order creating a voluntary framework for developers to provide the government access to covered frontier models for up to 30 days before release and to collaborate on selecting trusted partners for early access. The order also said nothing in that section should be read to create a mandatory licensing, preclearance or permitting requirement for new AI models.

Less than four weeks later, the distinction between voluntary coordination and government approval is already under pressure. OpenAI said Friday that it is beginning a limited preview of GPT-5.6 Sol, Terra and Luna, but that, at the government's request, access will start with a small group of trusted partners whose participation has been shared with the government. In its launch post, OpenAI said it does not believe this kind of government access process should become the long-term default because it keeps tools from users, developers, enterprises, cyber defenders and global partners.

OpenAI's move shows that Anthropic is no longer an isolated case. The government is now treating advanced cyber-capable models as systems whose release conditions can be negotiated before or after launch. The difference is procedural: OpenAI is cooperating before broad release, while Anthropic is negotiating after an abrupt shutdown.

The legal basis remains unsettled

The government's chosen instrument is also under scrutiny. A CSIS analysis said the June 12 action required a Bureau of Industry and Security license for foreign persons, inside or outside the US, to access Fable 5 and Mythos 5. CSIS also noted that the legal fit is uncertain because remote access to a model on Anthropic servers is not the same thing as exporting model weights or source code.

A Harvard Law Review essay framed the question more bluntly: when a user sends a prompt to a hosted AI model and receives a response, has the provider exported the model? That distinction matters because the government appears to be using export-control machinery built around goods, software, technology transfer and deemed exports to govern cloud access to a hosted AI service.

That uncertainty gives both sides an incentive to settle. For Anthropic, a negotiated lift would restore access without forcing an immediate courtroom test of Commerce's authority. For the administration, a deal would preserve the precedent that frontier model releases can be conditioned on government security concerns without risking an early legal ruling that narrows its power.

The unresolved question is what Anthropic must actually change. Bloomberg reported that restrictions could be lifted once officials across the administration approve, but the mechanics remain unclear. The likely options include additional access controls, monitoring, customer vetting, revised safeguards or some combination of those measures. What would matter most is whether any agreement becomes Anthropic-specific, or whether it becomes the playbook every frontier lab must follow before shipping a model with advanced cybersecurity capability.

For Brown, Amodei and Anthropic, the near-term objective is simple: get Fable 5 and Mythos 5 back online. The larger consequence is less contained. A company built around the premise that safety can be a competitive advantage is now negotiating with a government that may decide safety claims are not enough unless Washington also signs off on access.