Fable 5: The AI That Was Too Powerful for the Government — And Why It's Back

Three days. That's how long Claude Fable 5 was publicly available before the US government stepped in.

On June 9, 2026, Anthropic released its most capable AI model to the public. By June 12, export controls forced the company to suspend global access after Amazon researchers discovered a bypass technique that could prompt the model to produce exploit code for software vulnerabilities. For three weeks, the most powerful publicly available AI model in the world simply vanished.

On July 1, it came back. And the story of what happened in between tells you everything you need to know about where AI is heading.

What Makes Fable 5 Worth Banning

To understand why the government reacted the way it did, you need to understand what Fable 5 actually is.

Claude Fable 5 is the first "Mythos-class" AI model Anthropic has made publicly available. Mythos is the company's highest-tier model family, previously restricted to vetted government and enterprise partners. Fable 5 gives the public access to that level of capability — with safety classifiers layered on top.

The numbers are staggering. Fable 5 scores 80.3% on SWE-Bench Pro, the gold-standard benchmark for real-world software engineering. That's an 11-point jump over the previous flagship, Claude Opus 4.8, and a 22-point lead over OpenAI's GPT-5.5. On the Artificial Analysis Intelligence Index, it ranks first globally with a score of 65, beating both GPT-5.5 and Google's Gemini 3.1 Pro.

But benchmarks only tell half the story. Stripe used Fable 5 to migrate a 50-million-line Ruby codebase in a single day — work that would have taken a full engineering team over two months. Hex, the data analytics platform, saw a 10-point jump on their core analytical benchmark, breaking 90% for the first time.

Built for Marathon Work, Not Sprints

Previous AI models were essentially sophisticated question-answering machines. Fable 5 is something different: an agent that can be handed a specification and trusted to execute it over hours or even days.

The model ships with a one-million-token context window — matching Opus 4.8's ceiling but paired with far greater reasoning depth — and can output up to 128,000 tokens per request. It uses file-based persistent memory to build on prior state during long sessions. In controlled testing with the card game Slay the Spire, adding persistent memory improved Fable 5's performance three times more than it improved Opus 4.8, and the model reached the game's final act three times more often.

This isn't just a coding tool either. Fable 5 can extract precise numbers from scientific figures, reconstruct web apps from screenshots alone, and even beat Pokémon FireRed using nothing but visual input — where previous Claude models struggled even with additional helper tools.

The Three-Week Ban

So what actually happened?

Amazon researchers found that a specific prompting technique could push Fable 5 past its safety boundaries, generating code that exploited software vulnerabilities. The US government responded by applying export controls, requiring Anthropic to restrict access to foreign nationals. Since real-time nationality verification wasn't feasible, Anthropic had no choice but to pull the model entirely.

Anthropic pushed back, pointing out that GPT-5.5 and Claude Opus 4.8 could produce identical outputs under similar conditions. This wasn't a Fable-5-specific problem — it was a frontier-AI problem that Fable 5's power simply made more visible.

What followed was two weeks of intensive collaboration between Anthropic and the Department of Commerce's Center for AI Standards and Innovation. The result: an improved classifier system that blocks the bypass technique in over 99% of cases. Export controls were lifted on June 30, and Fable 5 returned globally on July 1.

The trade-off? The new classifier increases false positives on routine coding tasks — a deliberate "safety margin" that some developers will find annoying but that regulators considered necessary.

The Two-Model Safety Architecture

The ban and its resolution highlighted something genuinely novel about how Anthropic approaches safety with its most powerful models.

Fable 5 runs with safety classifiers that monitor requests in real time. When a request touches high-risk domains — cybersecurity, chemistry, biology, or model distillation — and triggers the classifier, it is automatically rerouted to the more constrained Claude Opus 4.8 at no additional charge. The API returns a refusal reason rather than an HTTP error, so developers can handle it gracefully.

Meanwhile, Claude Mythos 5 — the same architecture without the safety classifiers — remains restricted to approved partners through Anthropic's Project Glasswing program. It's the same raw power, but only for vetted research and cybersecurity teams.

Over 1,000 hours of external red-teaming found no universal jailbreaks for Fable 5's classifier system. Whether that holds as millions of developers push the boundaries remains the open question.

What This Means Going Forward

Fable 5's brief disappearance exposed a tension that isn't going away: AI models are advancing faster than the regulatory frameworks designed to govern them. Anthropic's solution — safety classifiers, fallback routing, and rapid collaboration with regulators — is pragmatic, but it's also a reminder that the gap between "powerful enough to be transformative" and "powerful enough to be dangerous" is vanishingly small.

For developers and businesses, the practical takeaway is clear. Fable 5 is available now through the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and Claude.ai. It costs $10 per million input tokens and $50 per million output tokens — double the rates for Opus 4.8, but the model tends to complete complex tasks in fewer steps, often making it more cost-effective in practice.

The AI that was too powerful for the government is back. It's a little more cautious now — but no less capable.

---

Sources: Anthropic, NBC News, Anthropic — Redeploying Fable 5, TrueFoundry, 9to5Google