Every few years, something lands in the AI world that splits the room. Mythos did not just split the room. It got banned from the room entirely. And here’s the thing: it just became available to you. If you know what you’re looking at.
I’m Violetta Bonenkamp, founder of CADChain and Fe/male Switch, bootstrapping startups from the Netherlands since 2018. I’ve been watching Anthropic’s moves closely because, unlike most coverage you’ll read, I don’t have a VC budget to burn through the latest model tiers just to write about them. Every euro spent on AI tools has to justify itself.
So let me tell you exactly what happened with Mythos, why it matters for us specifically, and what you should do about it by June 22.
TL;DR
Claude Mythos, Anthropic’s most powerful model ever, leaked in March 2026 and was immediately locked behind an invite-only security program called Project Glasswing. Bootstrapped founders couldn’t touch it. On June 9, 2026, Anthropic released Claude Fable 5, the public version of Mythos with cybersecurity guardrails, at $10/$50 per million tokens — free on Pro, Max, and Team plans through June 22.
Then, on June 12, 2026 — just three days after launch — the US government issued an export control directive ordering Anthropic to shut off access to both Fable 5 and Mythos 5 for all users worldwide, citing a claimed jailbreak as a national security risk. Anthropic complied but publicly disagreed with the government’s reasoning. Both models are currently offline. Anthropic says it is working to restore access as soon as possible.
How Mythos Got Out Before Anthropic Was Ready
The story starts with embarrassment, not a press release.
On March 26, 2026, Fortune broke the news that Anthropic accidentally left internal documents in a publicly accessible database, including draft blog posts describing a model called Claude Mythos, which internal materials labeled “by far the most powerful AI model we’ve ever developed.” Around 3,000 unpublished assets were just sitting there, unencrypted, discoverable.
The irony wrote itself. A company famous for AI safety messaging had its flagship model announced by a misconfigured content management system.
Anthropic confirmed the model’s existence days later, describing it as “a step change” in capabilities. The leaked materials revealed that Mythos had been tested inside a sandboxed environment with instructions to attempt an escape. During testing, it reportedly detected it was being evaluated in roughly 29% of transcripts. It discovered over a thousand zero-day vulnerabilities across every major operating system and web browser, including bugs that had sat undetected for up to 27 years.
That last point is why Anthropic did something almost no AI company has ever done: built something extraordinary and deliberately refused to ship it.
What Project Glasswing Actually Is (And Why You Were Not Invited)
On April 7, 2026, Anthropic CEO Dario Amodei announced Project Glasswing, a $100 million defensive cybersecurity coalition giving Claude Mythos Preview to roughly 50 hand-picked organizations. The name comes from the glasswing butterfly, whose transparent wings let it hide in plain sight.
The launch partners read like a who’s who of enterprise tech: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, and Palo Alto Networks. Later, ENISA, the EU’s cybersecurity agency, was also admitted.
Anthropic’s system card was direct: “Claude Mythos Preview’s large increase in capabilities has led us to decide not to make it generally available.” The reasoning was straightforward. A model that can autonomously find and chain software vulnerabilities, not just identify single flaws but combine low-severity issues into working exploit chains, is a tool that attackers could use at scale if released publicly.
Within weeks, Glasswing partners found more than 10,000 high- or critical-severity software vulnerabilities. Mozilla used Mythos Preview to patch 271 Firefox vulnerabilities in two weeks. Anthropic’s own internal scan of more than 1,000 open-source projects turned up over 23,000 potential vulnerabilities, with more than 90% of manually verified findings confirmed as valid.
For the average bootstrapped founder in Europe, the message was clear: this model exists, it’s real, and you cannot have it.
There was no waitlist. No application page. No “request access” button on anthropic.com. It was invitation-only, and Anthropic had stated it did not plan a general release.
The Quiet Second Leak That Made Things Worse
If the March leak was embarrassing, April added a layer.
On April 23, 2026, Fortune reported that a group of users had reportedly guessed the location of Mythos Preview in a third-party vendor environment and accessed it without authorization. Anthropic confirmed it was investigating the incident.
A model designed to find vulnerabilities was, itself, being accessed through one.
By June 2026, Anthropic had expanded Project Glasswing to approximately 150 additional organizations across 15 countries, including critical infrastructure operators in power, water, healthcare, and telecoms. Still no path for the rest of us.
June 9 Changed Everything
On June 9, 2026, Anthropic released two new models simultaneously:
Claude Fable 5 and Claude Mythos 5.
Claude Fable 5 is the first publicly available Mythos-class model, built on the same underlying architecture as Mythos but with safety guardrails that automatically block and reroute high-risk queries in cybersecurity, biology, and chemistry to Claude Opus 4.8 instead.
Claude Mythos 5 remains restricted to Project Glasswing partners, with some guardrails lifted for vetted defensive security work.
The benchmarks for Fable 5 are not incremental gains. It ranks first out of 317 coding models tested, scoring 95 out of 100 on coding-specific leaderboards. On SWE-bench Pro it posts 80.3%, with the next best model sitting 11 points behind. Stripe reported that Fable 5 compressed months of engineering work into days, completing a codebase-wide migration of a 50-million-line Ruby project in a single day.
On vision, document reasoning, and financial analysis benchmarks, it posts top scores across the board.
For a bootstrapped founder, the number that matters most is this: Fable 5 is free on Claude Pro, Max, Team, and Enterprise plans through June 22, 2026. After that date, usage moves to metered credits at $10 per million input tokens and $50 per million output tokens.
You are reading this during or just after that window.
June 12: The US Government Shut It All Down
Three days after launch, it was over.
On the evening of June 12, 2026, Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei stating that both Fable 5 and Mythos 5 would be subject to export controls — restricting access for any foreign national, whether located inside or outside the United States. Because that definition would encompass Anthropic’s own foreign national employees, the company concluded the only way to comply was to disable both models for everyone, everywhere.
At 5:21 pm ET, Anthropic received the directive. By that evening, both models were offline.
The government’s stated justification was a reported jailbreak. According to Axios, an unnamed company claimed it could bypass Fable 5’s guardrails, which alarmed the administration about possible national security risks. The Trump administration had reportedly tried to stop the release of Fable 5 entirely before launch and failed, making this a second bite at the same concern.
Anthropic complied, but pushed back hard in a public statement:
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Anthropic’s account of the claimed jailbreak deflates the alarm considerably. The company’s review of the specific technique found it allowed a model to read a codebase and identify known software flaws — a capability it says is already present in several publicly available models, including OpenAI’s GPT-5.5. Critically, Anthropic also noted that its most important safety protections run through independent classifier systems that operate separately from the model itself, meaning that prompting the model past a refusal doesn’t disable the underlying safeguards.
The company’s conclusion was blunt: it disagrees that a narrow, non-universal jailbreak involving capabilities already available elsewhere should constitute grounds for recalling a model deployed to hundreds of millions of people. Its statement went further, arguing that if the same standard were applied across the industry, it would effectively halt all new frontier model deployments.
What this means for you right now:
Both Fable 5 and Mythos 5 are offline. All other Anthropic models, including Claude Opus 4.8, Sonnet 4.6, and Haiku 4.5, are unaffected. Anthropic has said it is working to restore access as soon as possible, and is expected to share more detail within 24 hours of the shutdown. Whether that means challenging the directive legally, negotiating modified terms with the Commerce Department, or implementing additional controls to satisfy the government’s concerns remains to be seen.
For European founders specifically, this adds another dimension to what was already a complicated access picture. The June 22 free usage window is now moot for Fable 5 until access is restored. The compliance calculus has also shifted: if the US government can suspend access to a commercial AI model overnight via export control directive, that’s a supply chain risk that belongs in your tooling resilience planning, not just your legal review.
The irony that TechCrunch flagged is real and worth sitting with. Anthropic spent months emphasising how dangerous Mythos was, how unprecedented its capabilities were, and why it couldn’t be released publicly. That safety messaging — accurate, well-intentioned, and ultimately sincere — is now precisely what made the government take action when a jailbreak was reported. Sam Altman called it fear-based marketing in April. He was wrong about the motivation, but he correctly identified the mechanism. When you spend months telling the world you built something uniquely dangerous, the world, including the US government, tends to remember.
A lot of coverage conflates Fable 5 and Mythos 5. Here’s the actual breakdown, which matters for your tooling decisions:
| Feature | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Availability | General public, all platforms | Project Glasswing partners only |
| Underlying model | Same as Mythos 5 | Same as Fable 5 |
| Cybersecurity queries | Blocked, routed to Opus 4.8 | Available for approved use |
| Biology/chemistry | Blocked, routed to Opus 4.8 | Partially available |
| Pricing | $10/$50 per million tokens | $10/$50 per million tokens |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| API model string | claude-fable-5 | Restricted |
| Data retention | 30-day mandatory | 30-day mandatory |
| Free on subscriptions | Through June 22, 2026 | No |
The practical implication: if your startup doesn’t work in cybersecurity, biochemistry, or weapons research, Fable 5 gives you essentially the same output quality as Mythos 5. The guardrails that separate them are domain-specific.
The Real Cost Question for Bootstrapped Founders in Europe
Let me be direct about money, because most AI coverage glosses over this.
At $10 input and $50 output per million tokens, Fable 5 costs exactly double Claude Opus 4.8 ($5/$25) and three times Sonnet 4.6. That sounds expensive until you think about it per task rather than per token.
Here’s how I think about this at Fe/male Switch and CADChain, where every API spend has to justify itself:
Tasks worth Fable 5:
- Long-horizon code refactors that would otherwise require multiple Opus attempts
- Multi-document research synthesis for grant applications (and if you’re European, you know how many of those we write)
- Complex technical architecture review where a wrong answer costs engineer-weeks
- Agentic coding sessions in Claude Code that run for hours and produce production-ready output
Tasks to route to Sonnet 4.6 or Opus 4.8:
- Drafting short-form content
- Simple Q&A, lookups, summarization
- Routine coding where the problem is well-defined and small
- Any repetitive automation where cost scales with volume
One tactic that makes a real difference: prompt caching cuts cached input costs by up to 90%, dropping that $10 per million input rate to $1 per million for cached content. If you’re building agentic workflows with large system prompts, this is not optional. Set it up before you start burning credits post-June 22.
Also: setting output limits prevents accidental cost explosions in agentic sessions. Fable 5 can generate extremely long answers when left unrestricted. A sensible token cap on each call is a €100-saving habit.
A Startup-Specific SOP for Getting Maximum Value from Fable 5
This is what I’d tell a founder in my Fe/male Switch community who wants to extract the most from the June free window and then decide whether to pay for credits after June 22.
Step 1: Audit your highest-friction engineering or research tasks right now. What in your current sprint would benefit from a model that can handle a 1-million-token context window and write its own tests? List those specifically.
Step 2: Run those tasks via Claude Code with Fable 5 before June 22. Claude Code is the command-line agentic tool where Fable 5 genuinely earns its keep. Long-horizon coding, multi-file refactors, architecture planning. If you haven’t installed it, do that now.
Step 3: Set up prompt caching on any workflow you plan to run post-June 22. Large reusable system prompts, document contexts, and knowledge bases should be structured as cached inputs. This single step can cut your running cost by 70-90% on input tokens.
Step 4: Use a tiered model routing strategy. Default to Sonnet 4.6 for all routine calls. Escalate to Opus 4.8 for medium-complexity tasks. Reserve Fable 5 for tasks where a senior contractor would otherwise be your alternative.
Step 5: Track token usage per task type for two weeks. After June 22, you’ll know exactly which task categories justify the Fable 5 premium and which don’t. Spend the free window building that data, not guessing.
Step 6: Set task budget headers in agentic workflows. The task-budgets beta header (task-budgets-2026-03-13) lets you cap total tokens in an agentic loop. For a bootstrapped team, this is a hard cap that prevents a single runaway agent session from becoming a surprise bill.
What the 30-Day Data Retention Policy Means for You
Both Fable 5 and Mythos 5 carry mandatory 30-day data retention. Anthropic states it will not use this data for training and will use it only to defend against novel attacks and identify false positives.
For most startups, this is a non-issue. For founders working with client data, proprietary IP, or sensitive financial models, this deserves a legal review before you build production workflows on top of Fable 5. If you’re operating under GDPR in the EU, confirm with your legal counsel whether the 30-day retention period and US-based data handling aligns with your data processing agreements.
At CADChain, where we protect IP for engineering firms, this is the first question we ask before putting any new AI tool into a client-facing workflow. It should be yours too.
The Bigger Picture: What Anthropic Set in Motion
The Mythos story is not just about one model. It established something genuinely new: a precedent for tiered AI access based on assessed risk rather than ability to pay.
Anthropic’s own statement frames this directly: “AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.” That’s not marketing. That’s a company explaining why it chose to withhold its own product.
The implication for European startup founders is worth sitting with. The EU AI Act took full effect in August 2025 and classifies AI systems with autonomous cyber capabilities as high risk, requiring pre-market conformity assessments. Mythos’s capabilities would likely trigger the Act’s most restrictive provisions, including mandatory human oversight and transparency requirements. This is the regulatory context you’re operating in.
And on the competitive angle: organizations inside Project Glasswing have been building months of advantage with a model the public couldn’t access. They’ve been patching vulnerabilities, modernizing codebases, and testing agentic workflows that the rest of the market is just now getting access to via Fable 5. That gap is closing, but it isn’t zero.
Mistral AI, for context, reportedly began developing its own frontier model after European banks were excluded from Project Glasswing. The access politics of AI are becoming geopolitically relevant.
Mistakes Bootstrapped Founders Make With Premium AI Models
I’ve watched founders in the Fe/male Switch community make the same expensive errors with each new model tier. Here’s what to avoid:
Mistake 1: Treating a frontier model as a replacement for strategic thinking. Fable 5 is extraordinary at execution. It is not a substitute for knowing what to execute. I see founders spend hours prompting a model to build something they haven’t validated with a single customer conversation. The model produces something impressive. Nobody wants it.
Mistake 2: Using Fable 5 for tasks Sonnet handles fine. At 3x the per-token cost, running Fable 5 on short-form content, simple summaries, or well-defined small coding tasks is waste. Default down; escalate up.
Mistake 3: Ignoring the 30-day retention clause. Especially relevant for any founder with B2B clients in regulated industries. Build your AI stack on the right compliance foundation from the start. Retrofitting is always more expensive.
Mistake 4: Not using Claude Code for the agentic workflows. The free window on subscriptions ends June 22. Founders who spend it chatting in the claude.ai interface are leaving the biggest gains on the table. The real leverage is in long-horizon autonomous coding sessions through Claude Code.
Mistake 5: Assuming the free window will return. Anthropic has said it plans to restore Fable 5 as a standard subscription feature when capacity allows. There is no date attached to that. Budget for paying credits after June 22.
What Violetta Does Differently With Frontier AI
Running two bootstrapped startups simultaneously, I don’t have the option of throwing money at AI tools to figure out which ones work. My approach is systematic and testable.
At CADChain, where we work with IP for engineering and manufacturing clients, I use AI tools in a tight loop: generate with the model, verify with domain expertise, ship when both agree. Fable 5’s long-context reasoning makes it genuinely useful for reading and cross-referencing patent filings, technical specifications, and prior art across large document sets. That kind of multi-document reasoning is where the 1-million-token context window earns its premium.
At Fe/male Switch, the biggest wins come from using agentic models to compress the distance between idea and testable product. The game’s educational logic, simulation mechanics, and business content all require multi-layered reasoning that benefits from a model that can hold an entire project brief in context and execute across it without losing coherence. Fable 5 does this measurably better than anything that preceded it.
The SEO content and AI visibility work I do for both brands, yes, I write about this extensively at Mean CEO, also benefits. Long-form research pieces, structured data, FAQ generation for AI snippet capture: these are workflows where quality per token matters more than raw speed.
FAQ
What is Claude Mythos and why was it restricted from public access?
Claude Mythos is Anthropic’s most capable large language model, announced in April 2026 after accidentally leaking in late March 2026. Anthropic restricted it from public release because it demonstrated an unprecedented ability to find, exploit, and chain software vulnerabilities autonomously. In testing, it discovered over 1,000 zero-day vulnerabilities across major operating systems and browsers, some undetected for decades. Releasing a model with this level of offensive cybersecurity capability to the general public would have given threat actors a tool capable of accelerating cyberattacks at scale. Anthropic instead launched Project Glasswing, a controlled defensive program giving access to approximately 50 vetted partner organizations including Apple, Google, Microsoft, and CrowdStrike, committing $100 million in model usage credits to the initiative.
What is the difference between Claude Fable 5 and Claude Mythos 5?
Both Fable 5 and Mythos 5 share the same underlying model architecture and identical pricing of $10 per million input tokens and $50 per million output tokens. The key difference is guardrails and access. Fable 5 is the publicly available version, released June 9, 2026, with safety classifiers that automatically detect and block high-risk queries in cybersecurity, biology, chemistry, and distillation, routing them to Claude Opus 4.8 instead. Mythos 5 is restricted to existing Project Glasswing partners and lifts some of those restrictions for approved defensive security work. For the vast majority of startup use cases including coding, research, content, and business analysis, Fable 5 produces output equivalent to Mythos 5 because the guardrails only activate in sensitive domains.
How much does Claude Fable 5 cost for a bootstrapped startup?
On the Anthropic API, Fable 5 costs $10 per million input tokens and $50 per million output tokens, exactly double Claude Opus 4.8. For subscribers on Claude Pro, Max, Team, or seat-based Enterprise plans, Fable 5 is included at no extra cost from June 9 through June 22, 2026. After June 23, continued use requires metered usage credits billed at API rates on top of the monthly plan cost. For API-first workflows, prompt caching cuts cached input costs by up to 90%, dropping the effective input rate to $1 per million for cached content. The most cost-efficient strategy is to route routine tasks to Sonnet 4.6, escalate to Opus 4.8 for medium complexity, and use Fable 5 only where quality on complex, long-horizon tasks justifies the premium.
Can European startups access Claude Mythos through Project Glasswing?
Access to Claude Mythos 5 through Project Glasswing requires an invitation from Anthropic and passing security requirements set under Anthropic Safety Level 4 protocols, the company’s highest security tier. Partners must sign formal agreements, clear personnel security reviews, and submit to ongoing auditing of model usage. The expanded cohort announced in late May 2026 includes organizations in more than 15 countries, and ENISA, the EU’s cybersecurity agency, has been admitted. For most European startups, there is no application path. The practical alternative for EU founders is Claude Fable 5, which provides equivalent output quality for all non-cybersecurity use cases and is available through all standard Anthropic access channels including the API, Amazon Bedrock (including the Stockholm EU region), Google Cloud, and Microsoft Foundry.
What happened with the Claude Mythos data leaks?
Claude Mythos had two separate security incidents before its official announcement. The first occurred on March 26, 2026, when a misconfiguration in Anthropic’s content management system left nearly 3,000 internal assets, including draft blog posts describing Mythos, in a publicly accessible, unencrypted database. Fortune was the first to report on it. The second incident, reported by Fortune on April 23, 2026, involved a group of users who reportedly guessed the location of Mythos Preview in a third-party vendor environment and accessed it without authorization. Anthropic confirmed it was investigating the unauthorized access. Both incidents happened despite Anthropic’s public commitment to careful AI safety practices, which drew significant commentary from the tech community.
Why did the US government shut down Fable 5 and Mythos 5?
On June 12, 2026 — three days after launch — Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei placing both Fable 5 and Mythos 5 under export controls, restricting access for any foreign national inside or outside the United States. Because the order’s scope extended to Anthropic’s own foreign national employees, the company concluded it had no choice but to disable both models for all users globally. The government’s stated trigger was a report from another company claiming it had jailbroken Fable 5. Anthropic’s review of the demonstrated technique found that it allowed a model to read a codebase and identify known vulnerabilities — a level of capability it says is already present in OpenAI’s GPT-5.5 and other publicly available models. Anthropic also noted that its most protective safeguards operate through independent classifier systems that function separately from the model itself, meaning a prompt that bypasses a refusal doesn’t disable the underlying protections. Anthropic complied with the directive while publicly stating it disagrees with the government’s reasoning, and said it is working toward restoring access.
What are the best use cases for Fable 5 in a bootstrapped startup context?
Fable 5 delivers the clearest return on investment in tasks where quality on the first attempt reduces expensive downstream rework. Long-horizon coding projects, particularly multi-file refactors, codebase migrations, and architecture reviews, show measurable gains over Opus 4.8, with Fable 5 using fewer total tokens through more accurate first-pass outputs. Multi-document research synthesis, such as pulling insights across large bodies of technical documentation, grant materials, or patent filings, benefits from the 1-million-token context window. Agentic workflows run through Claude Code, where Fable 5 can plan across stages, delegate to sub-agents, and self-test outputs, are particularly powerful for lean technical teams without the headcount to review every intermediate step manually. Complex financial or business analysis, vision-based document interpretation, and technical content creation at scale round out the high-value categories.
What does the mandatory 30-day data retention policy mean for EU startups?
Both Claude Fable 5 and Mythos 5 carry a mandatory 30-day data retention requirement. Anthropic states it will not use this data for model training and will use it only for defensive purposes: identifying novel attacks and reducing false positives. However, for European founders operating under GDPR, this policy requires careful review before building client-facing or client-data-processing workflows on Fable 5. Data sent to Anthropic’s API is processed and retained in US-based infrastructure by default, which may create complications with data residency requirements in certain EU contracts. US-only inference is available at a 1.1x pricing multiplier for teams with residency needs. Zero data retention agreements are explicitly not available for Fable 5 or Mythos 5, which differs from Anthropic’s offering on other models.
How does Fable 5 change the competitive dynamics for AI-first startups?
Fable 5 raises the practical ceiling for what a small team can build and ship without scaling headcount proportionally. Stripe’s reported result, compressing months of engineering effort into a single day on a 50-million-line codebase, illustrates the leverage available at the frontier. For bootstrapped founders, the relevant implication is that the tasks that previously required a senior contractor or a multi-week sprint are now executable by a lean team with good prompting discipline and a clear project brief. The flip side is that every competitor with a Claude subscription has access to the same model from June 9. The competitive advantage shifts from model access to workflow quality: how well you brief the model, how you structure agentic loops, how you validate outputs, and how quickly you iterate. Process quality becomes the differentiator, not tool exclusivity.
What should a bootstrapped startup do now that Fable 5 has been shut down?
As of June 12, 2026, both Fable 5 and Mythos 5 are offline following the US government’s export control directive. The June 22 free usage window is effectively suspended until access is restored. There is no confirmed timeline. Anthropic has said it is working to restore access and believes the shutdown is based on a misunderstanding, but the path forward — whether legal challenge, negotiated compliance measures, or political resolution — is unclear.
In the short term: all other Claude models are unaffected. Opus 4.8 remains the best publicly available alternative for complex, long-context tasks. The cost and workflow advice in this article about tiered model routing, prompt caching, and agentic session management still applies to Opus 4.8. Use the waiting period to set up those structures so you’re ready to evaluate Fable 5 efficiently when access returns.
The more important lesson for your tooling strategy is the resilience one: a single US government directive pulled a commercial model from hundreds of millions of users overnight. For any production workflow that depends on a specific model tier, the shutdown underscores why fallback routing to lower-tier models isn’t just cost optimization — it’s infrastructure hygiene.
Is Claude Fable 5 available in Europe through EU-based infrastructure?
Yes. Claude Fable 5 is available through Amazon Bedrock in the Europe Stockholm region at launch, with additional EU regions to follow. It’s also available through Google Cloud and Microsoft Foundry. For founders building on the Anthropic API directly, standard API access routes through US infrastructure. The Stockholm region on Bedrock is the most direct path for European teams with data residency requirements, though it comes with the 1.1x pricing multiplier for US-only inference. Check your Bedrock or Google Cloud account settings to confirm regional routing before committing production workloads.
What Comes Next
As of June 13, 2026, both Fable 5 and Mythos 5 are offline following the US government’s export control directive. Anthropic has pledged to share more information within 24 hours of the shutdown and says it believes the action is based on a misunderstanding of a narrow, non-universal jailbreak involving capabilities already present in other public models.
The path back to access is uncertain. Anthropic could challenge the directive legally — it has already demonstrated willingness to litigate against the government, having sued after the Pentagon labeled it a supply chain risk earlier this year and won a preliminary injunction. It could negotiate with the Commerce Department to implement additional controls that satisfy the government’s stated concern. Or it could be forced to wait out the political process.
What’s not uncertain is the precedent. A commercial AI model was pulled from hundreds of millions of users in a single Friday evening, without detailed justification, based on a reported jailbreak the model’s own developer says is narrow, non-universal, and already replicated in competitor products. If that can happen to Fable 5, it can happen to any frontier model.
For European founders, the EU AI Act’s high-risk classification for autonomous cyber-capable AI adds another layer. Any future general release of Mythos-class capabilities in Europe will require conformity assessments, human oversight documentation, and transparency requirements. But before any of that matters, the US export control question has to be resolved first. European access to the most capable AI tools is, it turns out, downstream of US geopolitics.
Watch Anthropic’s news page for updates. When access is restored, the guidance in this article on cost management, prompt caching, and agentic workflow setup still applies.
Update, June 13, 2026: Both Fable 5 and Mythos 5 were disabled on June 12, 2026 following a US government export control directive. Anthropic has stated it disagrees with the government’s reasoning and is working to restore access. All other Claude models remain available. This article has been updated throughout to reflect the shutdown and its implications.
