โ๏ธArs Technica AIโขStalecollected in 30m
Anthropic Shuts Down Fable and Mythos Models
๐กFirst major model shutdown due to US government national security directives; a critical precedent for AI safety.
โก 30-Second TL;DR
What Changed
Anthropic terminated Fable and Mythos models due to government intervention.
Why It Matters
This move signals a tightening regulatory environment for AI developers, where government oversight can lead to the immediate shutdown of deployed models. Companies must now prioritize robust safety and jailbreak prevention to avoid similar intervention.
What To Do Next
Review your model's safety guardrails and red-teaming protocols to ensure they meet emerging national security compliance standards.
Who should care:Founders & Product Leaders
Key Points
- โขAnthropic terminated Fable and Mythos models due to government intervention.
- โขThe U.S. Department of Commerce cited national security risks.
- โขSpecific concerns were raised regarding jailbreak vulnerabilities in Fable 5.
๐ง Deep Insight
Web-grounded analysis with 22 cited sources.
๐ Enhanced Key Takeaways
- โขThe U.S. Department of Commerce's directive was an export control order, specifically mandating the suspension of Fable 5 and Mythos 5 access for any foreign national, including Anthropic's own foreign-national employees, which led Anthropic to disable the models for all users to ensure compliance.
- โขAnthropic publicly disputed the severity of the alleged jailbreak vulnerability in Fable 5, stating that the technique appeared narrow, not universal, involved previously known minor vulnerabilities, and that their extensive red-teaming efforts had not uncovered a universal jailbreak.
- โขClaude Fable 5 was the publicly available 'Mythos-class' model, while Claude Mythos 5 was a less restricted version with stronger cybersecurity capabilities, initially deployed through an invitation-only 'Project Glasswing' for trusted partners, including the US government.
- โขThe government's intervention is linked to a recent Executive Order issued by President Trump on June 2, 2026, titled 'Promoting Advanced Artificial Intelligence Innovation and Security,' which calls for a framework for developers to engage with the federal government for pre-release review of frontier AI models due to national security and cybersecurity risks.
- โขAnthropic has a history of disagreements with the Trump administration over AI safety and defense applications, including a prior legal standoff and a Pentagon designation as a supply chain risk in February 2026 after the startup sought limits on its AI model for certain defense uses.
๐ Competitor Analysisโธ Show
| Feature/Model | Anthropic Fable 5 / Mythos 5 | OpenAI GPT-5.4 / GPT-5.3 Codex |
|---|---|---|
| Capability Class | Mythos-class (frontier) | Frontier |
| Primary Use Cases | Complex reasoning, multi-step agentic workflows, advanced code generation, research synthesis, cybersecurity vulnerability detection (Mythos 5) | General-purpose LLM, coding, reasoning |
| Context Window | 1 million tokens | (Not specified in search results for GPT-5.x) |
| Max Output Tokens | 128,000 tokens | (Not specified in search results for GPT-5.x) |
| Safety Measures | Constitutional AI, strong safeguards, fallbacks to Opus 4.8 for high-risk domains (Fable 5), 30-day data retention for misuse research | (General LLM safety measures) |
| Benchmarks (AISI 2026) | Mythos 5 ranked highest, ahead of GPT-5.4 and GPT-5.3 Codex | Ranked below Mythos 5 |
| Pricing (Input/Output per million tokens) | Fable 5 / Mythos 5: $10 / $50 | (Not specified in search results for GPT-5.x) |
๐ ๏ธ Technical Deep Dive
- Model Class: Fable 5 and Mythos 5 belong to Anthropic's 'Mythos-class,' which is designed for complex, multi-step, and cognitively demanding tasks, sitting above the Opus class in capability.
- Context Window: These models feature an extended 1-million-token context window, allowing for large-scale code analysis, lengthy document synthesis, and maintaining coherence across hundreds of tool calls.
- Output Tokens: They support a maximum of 128,000 output tokens.
- Safety Architecture (Fable 5): Fable 5 incorporates strong safety classifiers that automatically route requests in high-risk domains, such as cybersecurity and biology, to a less capable model like Claude Opus 4.8.
- Mythos 5 Distinction: Claude Mythos 5 is the same underlying model as Fable 5 but with these specific safeguards lifted, making it more capable in sensitive areas like vulnerability discovery and biodefense screening, and thus restricted to trusted partners.
- Training Philosophy: Anthropic employs 'Constitutional AI,' a technique to improve ethical and legal compliance and align AI behavior with human values.
- Data Retention Policy: For Mythos-class models, Anthropic requires a 30-day retention of customer data to enable research and mitigation of potential jailbreaks and patterns of misuse.
- Jailbreaking Vulnerabilities: Jailbreak prompts are inputs designed to bypass LLM safety controls, often by tricking the model into adopting alternate identities (role play), distracting it (attention shifting), gaining higher access (privilege escalation), or using obfuscated language.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
There will be increased government scrutiny and potentially more stringent regulatory frameworks for frontier AI models globally.
The U.S. Department of Commerce's directive sets a precedent for government intervention based on national security concerns over potential misuse, even if disputed by the developer, indicating a more active regulatory stance.
AI developers will likely implement more robust pre-release safety evaluations and enhance collaboration with government agencies.
The recent Executive Order encourages a voluntary framework for pre-release review, and this incident highlights the significant consequences of not aligning with government safety expectations, pushing companies towards proactive engagement.
The AI industry will see an accelerated focus on developing and implementing more sophisticated and verifiable jailbreak resistance mechanisms.
The incident underscores the critical importance of preventing model misuse through jailbreaks, prompting a push for more advanced and transparent safety measures across all AI providers.
โณ Timeline
2021-01
Anthropic PBC founded by former OpenAI members.
2023-03
Anthropic releases its first public AI model, Claude.
2024-03
Anthropic launches Claude 3 models (Opus, Sonnet, Haiku).
2026-02
Pentagon designates Anthropic as a supply chain risk amid ongoing disputes over AI safety and defense applications.
2026-06-02
President Trump issues Executive Order 'Promoting Advanced Artificial Intelligence Innovation and Security'.
2026-06-09
Anthropic officially launches Claude Fable 5 and Claude Mythos 5.
2026-06-13
U.S. Department of Commerce issues directive, leading Anthropic to shut down Fable 5 and Mythos 5.
๐ Sources (22)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Ars Technica AI โ