The strongest Claude, stuck on the eve of going public

The most powerful model of Anthropic was suspended by the US government just three days after its launch.

On June 12, Anthropic issued an announcement stating that the US government, citing national security reasons, issued an export control directive requiring the suspension of any foreign nationals' access to Claude Fable5 and Claude Mythos5. Whether a person is inside or outside the US, even including foreign employees within Anthropic, are all included in the scope of restrictions.

To ensure compliance, Anthropic ultimately chose to temporarily disable these two models for all customers. Other Claude models are not affected.

This is not an ordinary model removal.

What it truly reveals is the new situation of cutting - edge AI companies: the stronger the model, the greater the commercial value, and the higher the regulatory uncertainty. In the past, US AI control mainly targeted chips, tools, and computing power resources. Now, regulation is starting to directly reach for model access rights.

For Anthropic, which is preparing to go public, Fable5 was originally a key card to prove its technological leadership and accelerated commercialization. However, just as the card was played, the regulators took their seats at the table first.

01: The most powerful model hits sensitive capabilities first

Fable5 and Mythos5 are not an ordinary upgrade.

According to Anthropic's official documentation, Fable5 is currently Anthropic's "most powerful widely - released model", designed for high - difficulty reasoning and long - cycle intelligent agent tasks. It has been available since June 9 through channels such as Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry, supporting a 1 - million - token context, priced at $10 per million input tokens and $50 per million output tokens.

This is no longer just a model that is "better at chatting".

Anthropic said in the release that Fable5 and Mythos5 can work autonomously for a longer time than previous Claude models and are suitable for scenarios such as software engineering, knowledge work, vision, memory, and life science research. In other words, it aims to solve not simple Q&A but more complex, longer - chained tasks closer to real - world production systems.

This is where the problem lies.

The stronger the model, the more it has a dual nature. It can help programmers write code, help enterprises handle knowledge work, and help researchers propose hypotheses. It can also help security teams discover vulnerabilities and patch systems. However, the same capabilities can also be used by attackers to find system weaknesses, generate attack paths, and improve the efficiency of cyberattacks.

Mythos5 is more sensitive than Fable5.

Anthropic said that Mythos5 and Fable5 are based on the same underlying model, but the security guards are removed in some areas. Initially, it was only available to a small number of cyber defenders and infrastructure providers through Project Glasswing. Anthropic even directly stated that Mythos5 has "the world's strongest cybersecurity capabilities".

This statement is a technological selling point at a commercial press conference, but it is a risk signal in the eyes of regulatory authorities.

Project Glasswing has proven the real - world value of this capability. Anthropic said that initially about 50 partners used Claude Mythos Preview to scan code libraries and have discovered more than 10,000 high - risk or critical - level security vulnerabilities. On June 2, Anthropic announced the expansion of the project to about 150 new institutions, covering more than 15 countries, including key infrastructure industries such as power, water, healthcare, communications, and hardware.

This shows that AI has begun to enter the core of cybersecurity.

However, cybersecurity is naturally dual - purpose for offense and defense. The more a model can help the defense side discover vulnerabilities, the more the regulators may worry: if it falls into the wrong hands, will it also help the attackers improve efficiency?

Therefore, what the US government is targeting this time is not an ordinary chatbot, but a cutting - edge model that has entered the sensitive area of cyber offense and defense.

The stronger the model, the closer the regulation.

02: The stronger Anthropic is, the harder it is to go public lightly

What is really hit by this incident is Anthropic's narrative for going public.

On June 1, Anthropic secretly submitted a Form S - 1 registration draft to the US SEC in preparation for an IPO. On June 9, it released Fable5 and Mythos5, bringing the fifth - generation models to the market. On June 12, the US government required the suspension of foreign nationals' access to these two models.

In just over a dozen days, three actions are linked together: sprinting for the capital market, releasing the most powerful models, and facing export controls.

This has complicated Anthropic's story.

In the eyes of the capital market, Anthropic originally needed to prove that it is not just a follower of OpenAI, but an AI company that can continuously launch cutting - edge models, attract enterprise customers, and integrate into cloud platforms and developers' workflows.

Fable5 is the key card in this story.

It is not only an upgrade in capabilities but also a commercial upgrade. The pricing of Fable5 and Mythos5 is higher than most general models. They are connected to mainstream cloud platforms and developer channels and attempt to promote scenarios such as "long - cycle intelligent agent tasks", "complex coding", and "enterprise knowledge work" that are more likely to generate payments.

What Anthropic really wants to tell is that the model is not just about answering questions, but can take on tasks, run operations, and deliver results.

However, the US government's directive has introduced another variable.

Cutting - edge models are not ordinary SaaS products. For ordinary SaaS products, as long as they can be sold, renewed, and expanded, a growth story can be told. Cutting - edge AI models also need to answer a more difficult question: who can use them, where can they be used, to what extent, and whether they will cross the national security boundary.

The concerns of enterprise customers have actually emerged first.

Microsoft's official documentation shows that Anthropic's preview models with data retention requirements, such as Claude Fable5 and Claude Mythos5, are disabled by default in Microsoft Online Services and need to be enabled separately by the tenant administrator. After enabling, Anthropic will retain most inputs and outputs for up to 30 days. If the content is identified by the security classifier as possibly violating the usage policy, the relevant content can be retained for up to 2 years, and the classification scores can be retained for up to 7 years.

This means that when the most powerful models enter the enterprise scenario, they not only need to compete in capabilities but also need to undergo customers' data compliance reviews.

For Anthropic, the problem is not that the models are not strong, but that the stronger the models are, the more customers and regulators will ask: How should the data be retained? How should the risks be managed? How should the access rights be defined? Who should be responsible if something goes wrong?

The US government's directive this time has magnified this uncertainty to the national security level.

Anthropic does not agree with the government's judgment. The company's announcement states that it received the government's directive at 17:21 Eastern Time on June 12, but the letter did not specify the specific national security concerns. Anthropic understands that the government may think that there is a "jailbreak" way to bypass the security guards in Fable5, which can be used to identify software vulnerabilities. Anthropic refuted that the demonstrations it has seen only involve a small number of known and minor vulnerabilities, and other public models can also find similar problems without bypassing the protection.

Anthropic also emphasized that Fable5 has undergone thousands of hours of red - team testing by the US government, the UK AI Safety Institute, third - party institutions, and its internal team before its launch. Fable5 itself has strong security guards. Once a request involves high - risk areas such as cybersecurity, biology, and chemistry, it will be transferred to Claude Opus4.8 for processing.

However, the regulatory logic does not only look at "whether there have been accidents".

Regulators are more concerned about: if this capability is used by people who should not have it, will it magnify the risks of cyberattacks, biosecurity, and critical infrastructure?

This is the current difficulty for Anthropic.

The more it tries to prove that the model is strong, the more it proves that there is a reason for regulation to intervene. The more it tries to expand commercialization, the more it has to accept the re - definition of access rights.

03: The moat of cutting - edge AI companies now includes regulatory credibility

The real industry signal of this incident is that the boundaries of US AI control are expanding.

In the past, the most core constraints in AI competition were at the underlying level. Whoever could obtain advanced GPUs, sufficient computing power, and access to cloud infrastructure had a better chance of training leading models. Therefore, in the past few years, the focus of US export controls has mainly been on AI chips and related tools.

However, the suspension of access to Fable5 and Mythos5 shows that regulation is moving up to the upper level.

After the model is trained, can it be launched globally? Can foreign nationals call it? Can it be distributed through cloud platforms and developer tools? Can enterprise customers access it without discrimination? These issues will become new regulatory variables.

This is a reminder to the entire cutting - edge AI industry.

AI companies used to like to talk about "global availability". Once a model is launched, it expands globally through APIs, cloud markets, enterprise services, and developer tools. This logic is very Internet - like and appealing to the capital market: low marginal cost, fast distribution, more customers, and higher revenue.

However, cutting - edge models are becoming less and less like ordinary Internet products.

They are more like chips, satellites, cryptographic systems, and high - end manufacturing equipment. The greater the commercial value, the stronger the strategic attribute; the higher the replicability, the more justifiable the regulatory intervention.

This is why "a strong model" does not make regulation disappear but rather attracts it.

The commercial value and security risks of a strong model come from the same source of capabilities. The more it can write code, find vulnerabilities, execute long - term tasks, and enter the enterprise production system, the more it may change the efficiency structure of cyber offense and defense.

For Anthropic, this suspension may not destroy its commercialization, but it will change the way the market prices it.

Investors cannot only look at model rankings, customer growth, and API revenue but also need to look at regulatory uncertainty. A cutting - edge AI model company in the future not only needs to prove its technological leadership but also needs to prove its ability to handle the complex relationships between the government, cloud providers, enterprise customers, and security institutions.

What is most worthy of attention about Anthropic this time is not whether it has been "banned".

To be precise, other Claude models are not affected, and Fable5 and Mythos5 were not removed because they have been confirmed to cause major security accidents. What it really exposes is that the moat of cutting - edge AI companies is no longer just model capabilities but the ability to balance capabilities, distribution, and compliance.

This is the essence of Anthropic being put on hold.

It is not that it has suddenly become weak, but that it has become so strong that it has entered another rule system.

In the past, large - model companies competed on who was smarter. Now, they also need to compete on another issue: who can legally, stably, and sustainably deliver the most powerful capabilities to customers.

What Anthropic needs to sell before going public is not just a stronger Claude but a set of delivery capabilities that can be accepted by regulators, customers, and the capital market at the same time.

This is the new threshold for the next round of cutting - edge AI companies.