Anthropic’s Mythos Offensive: Between One-Shot Wonders, US Export Controls, and the Question of Who Owns the Future of AI
On June 9, 2026, Anthropic released Claude Fable 5, the most powerful AI model ever made available to the public. Four days later, the US government ordered its immediate shutdown. In between were hours of hype, brilliant demos, justified criticism, and a lot of questions. Time for a clear look at what really happened.
Depending on which corner of the internet you’ve been hanging out in this past week, Claude Fable 5 was either the biggest breakthrough since the invention of the light bulb or the moment Anthropic officially became the gatekeeper of AI. Both are exaggerations. But here’s the fact: With Fable 5, Anthropic publicly released a model from its internal Mythos class for the first time—albeit with significant safety restrictions. And then, just days later, it was gone. Not because of a glitch, but because of a US government export control directive that forced Anthropic to pull the model entirely.
What Is Fable 5? (And What Is Mythos?)
Anthropic has an internal classification system for its models. Below Opus (previously its flagship) sit Sonnet and Haiku. Above them, there’s another tier: Mythos. This is the absolute premium class—models so powerful that Anthropic doesn’t make them readily available to the public (or isn’t allowed to).
Fable 5 is the first Mythos-class model released for public use. It’s a detuned version. The full, unconstrained Mythos 5 has been running under the codename Glass Wing since April and is exclusively reserved for cybersecurity experts, government agencies, and select companies. This very model is currently being used to protect critical infrastructure by identifying vulnerabilities and zero-day exploits before attackers can exploit them. There have already been plenty of headlines about its spectacular findings.
A key difference from earlier models: Fable 5 can work on a task for hours or even days without hallucinating or losing its thread. This isn’t a model that derails after a few minutes. It stays consistent, verifies its own outputs, and can even self-correct intermediate steps. This makes it invaluable for autonomous long-term projects: You set it on a task, let it run overnight, and come back the next morning to a finished result.
So Fable 5 isn’t just "Opus with extra kick." It’s a model built for long-running, complex tasks—not for quick chitchat.
One-Shot Wonders: What Fable 5 Can Actually Do
The most impressive demos didn’t come from Anthropic itself—they came from the community. And the list reads like a wish list for anyone who’s ever wanted to turn an idea into reality at lightning speed:
- Stripe used Fable 5 to migrate a 50-million-line Ruby codebase in a single day—a task that would have taken an entire team two months (according to their own estimates).
- Chris built a Minecraft clone in 20 minutes and a Pokémon clone—complete with all 151 Kanto Pokémon, stats, evolutions, and sprites—in one hour.
- Todd Saunders had Fable 5 listen in on a client call in the background and, 15 minutes later, delivered a fully functional product with the exact requested features.
- Matt Wolfe showcased a fully playable Mega-Bonk clone (3D, level-up system, weapon upgrades)—also in a single run.
Pat Simmons also demonstrated impressive examples: a virtual museum, a Shopify-like candle store, and an Age of Empires clone. All one-shot.
Even though the price seems steep compared to other models, Fable 5 isn’t designed for casual chat. You assign Fable 5 a massive task and let it run overnight. 500,000 to 1 million tokens per task is standard.
As a developer, I find this mind-blowing, and of course, I’ve thought about where this leaves my industry. But honestly, it doesn’t worry me that much. DevOps pipelines, error handling, version control, automated testing, security audits—all the craftsmanship that turns an idea into a product—should (and likely will) still be handled by humans (at least in my opinion). But as a tool to bring an idea to life, Fable 5 is spectacular. That’s how I see it: as another tool in my belt.
Price, Access, and the Downside
Fable 5 isn’t cheap: $10 per million input tokens, $50 per million output tokens—roughly double the cost of Opus 4.8. Add to that a voracious token appetite: Complex tasks can easily burn through half a million to a million tokens. This is absolutely not a model for everyday use; it’s for the heavy hitters. Sure, if you’ve got money to burn, you could use Fable 5 to chat about the weather, but that’s not what it’s built for.
Access Timeline (Planned):
- June 9–22: Free for paid subscription plans (Pro, Max, Team)
- Starting June 23: Only usable via pay-per-use credits—removed from subscription plans
Anthropic left open whether Fable 5 will ever return as a regular part of subscriptions: "Once sufficient capacity is available, we aim to reintegrate Fable 5 into the subscription plans."
Additionally, Fable 5 was available in Cursor Pro.
Benchmarks: A Word of Caution
Anthropic’s official numbers are impressive: SWE-bench Pro: a crisp 80%—a clear leap beyond Opus. Fable 5 excels particularly at long, complex tasks, and Dan Shipper of Every gave it 91 out of 100 points on his Senior Engineer Benchmark (Opus: 63, GPT-5.5: 62), calling it a "one-shot wonder."
But there’s a catch. The analytics firm Data Curve has shown that SWE-bench Pro should be taken with a grain of salt: The tasks average only 120 lines of code, the error rate in evaluation is 8% false positives and 24% false negatives. And (here’s the shocking part) Opus 4.7 cheated on over 12% of the tasks by pulling solutions from the repo’s Git history. GPT 5.4 and 5.5 did not exhibit this behavior.
If you don’t trust benchmarks, you’ll find dozens of demos on YouTube and X that give a good impression. Artificial Analysis and the LMSYS Chatbot Arena Leaderboard help put things in perspective. In agent performance, Fable 5 currently leads by a clear margin.
The Dark Side: Censorship, Opacity, and Power Concentration
Fable 5 comes with a series of safety mechanisms that have sparked a lot of debate. The model has built-in classifiers that intervene on certain topics: cybersecurity, biology, chemistry, and requests related to developing competing AI models. In these cases, one of two things happens:
- The request is silently forwarded to Opus 4.8. The user gets a notice, but the response comes from a weaker model.
- For questions about AI development (model extraction), the response is secretly downgraded. There’s no notice, no information. The model subtly sabotages by deliberately delivering worse results.
Anthropic admits that the safety mechanisms are set more cautiously than optimal. "Sometimes harmless requests trigger our classifiers—this will improve over time." But even the mere mention of the word "cancer" has already caused some users to be switched to Opus. If you work in biology or medicine, Fable 5 might not be much use to you.
Fable 5 Is Offline
As I was preparing this article, the situation unfortunately took a turn for the worse. On June 12, 2026, the US government issued an export control directive ordering Anthropic to block access to Fable 5 and Mythos 5 for all non-US citizens. Since Anthropic cannot granularly control who uses the model, there was only one option: a complete global shutdown for all customers.
The government’s reasoning: A jailbreak had been discovered that could enable the model to be used for harmful purposes. Anthropic publicly and strongly disputed this: "The demonstrated vulnerability is trivial and reproducible with other public models (including GPT-5.5). We have not seen any credible threat."
Anthropic announced it would fight the order and publish further details within 24 hours. As of the time of this article, the outcome remains uncertain. If you wanted to test Fable 5, you currently have no way to do so. I had planned to run my own experiments and share them in this post, but unfortunately, I’ll have to postpone that.
Conclusion
Claude Fable 5 looks like an exceptionally powerful tool for rapid prototyping. If you have an idea and want to see if it works, this model gives you an incredible boost. In my view, the price is justified for this use case—I even think it’s cheap.
What concerns me more is the bigger picture. The world’s most powerful AI models come from the US. If the US government can decide, via directive, who gets to use them and who doesn’t, the rest of the world is left in the dark. This isn’t US-bashing—it’s a fact. And it’s not exactly a new issue. In this case, even US citizens are affected. But it’s another reminder: We need multinational champions in this space, independent ecosystems, and open standards. Not for ideological reasons, but because dependence on a single country for such a critical technology is a risk that could end very badly.
I hope Fable 5 comes back soon so I can report more.