Anthropic’s Fable 5 was the AI model everyone suddenly wished they could use. It outperformed OpenAI’s GPT-5.5 by a wide margin on every major benchmark. Then the U.S. government pulled the plug.

Fable 5 launched on June 9, 2026, promising to rewrite what AI could do in software engineering, reasoning, and complex multi-step tasks. It crushed GPT-5.5 on coding benchmarks like SWE-Bench Pro, scoring 80.3% to GPT-5.5’s 58.6%. That’s a leap, not a nudge. It even dominated agentic terminal tasks and long-horizon workflows where AI must think and act over hours.

Despite that, just three days later, an export control directive ordered Anthropic to shut down Fable 5 and its broader Mythos 5 family. The government cited a “jailbreak vulnerability” as the cause. Anthropic pushed back, calling the issues minor and publicly known, achievable by GPT-5.5 anyway. Still, the model vanished from public access.

This leaves GPT-5.5 the default champ by absence, not merit. It’s the best model you can use, not the best model that exists. For developers and researchers, that’s a bitter downgrade. Fable 5’s lead meant solving four out of five real-world coding problems, compared to GPT-5.5’s three. In high-stakes production environments, that difference matters.

Fable 5’s architecture was unique. It came with built-in safety classifiers that silently rerouted sensitive queries—cybersecurity, biochemistry, or attempts to extract the model’s capabilities—to a weaker sibling, Opus 4.8. These classifiers triggered under 5% of the time but prevented misuse. Anthropic’s compromise: top-tier power with guardrails.

On the price front, Fable 5 was no bargain. It charged twice GPT-5.5’s input token rate and nearly four times for outputs. For many teams, that’s a dealbreaker. But for those who could afford the premium, Fable 5’s ability to compress months of engineering work into days was a game changer. Stripe’s real-world test on a 50-million-line codebase migration proved it.

Anthropic’s split branding added complexity. Mythos 5 was the unrestricted version, limited to trusted partners. Fable 5 was the public face, deliberately capped for safety. No other AI lab has tried this “two doors, one brain” approach at the flagship level. It signals how frontier AI is moving toward controlled access for the riskiest capabilities.

Meanwhile, Opus 4.8 remains the practical choice for most. It outperforms GPT-5.5 on coding and knowledge tasks, costs less, and has a more forgiving safety policy. But it’s a notch below Fable 5, especially on multi-step agentic work where sustained reasoning counts.

Fable 5’s forced disappearance exposes a growing tension in AI. The fastest, smartest models raise new risks that governments want to control. Export rules, data retention requirements, and safety constraints clash with the open innovation ethos. The result: the best AI sometimes isn’t the one you get to use.

For now, Anthropic is negotiating with regulators. Whether Fable 5 returns or becomes an exclusive tool for select partners remains unclear. The AI landscape just got a little less exciting—for everyone but those who already had Mythos access.