In June, a US export-control order forced Anthropic to suspend Fable 5 and Mythos 5 for any foreign national. It did so within hours by switching them off worldwide. It couldn’t filter users by nationality in real time. The outage hit claude.ai, the API, Claude Code and Cowork at once.
Three days passed between a frontier model’s public launch and its government-ordered shutdown. We measure models on capability. What took two of them offline was governability: whether you’re allowed to run them. Almost no team puts a number on it.
You rent these models from a landlord who answers to a government. You do not hold the kill switch.
Three forces in one month
The order wasn’t alone.
In the same weeks, states from both parties passed AI laws despite a federal warning to hold off. That splits compliance into rules that apply according to where each user is, not where you’re headquartered.
And the copyright reckoning kept adding up. Anthropic’s $1.5 billion settlement over training data was still fresh. It’s among the largest copyright settlements in US history.
None of the three is about whether the model is good. Each can change what you can run, and for whom, overnight.
A reasoning score tells you nothing about the questions that matter here. Will you be allowed to serve a model to your users in six months? Will an export rule reach it? Will a court force a change to how it was trained? One already has, against another lab in Europe.
Your continuity isn’t fully in the lab’s hands either. When the order landed, Anthropic complied because it had to, while disagreeing in public. The decision was never its to make. And the most capable model in a lineup is the one most likely to draw the order that removes it.
Build as if nothing is permanent
Frontier models are the most valuable software most companies can adopt. Sitting them out would cost far more than the risk. The mistake is building as if any one model, lab or country were permanent.
The teams that absorbed the Fable shutdown in minutes had made model choice a setting, with a tested fallback behind it. Often that was Opus 4.8, which Fable was already routing some work to. The teams that scrambled had built one model in as load-bearing infrastructure.
So:
These now move your stack as much as the next benchmark does. Only the benchmark shows up on the leaderboard.
Frequently asked questions
What is political risk in an AI roadmap?
It's the chance that forces outside model quality, like export controls, regulation or lawsuits, change what you can run, for whom, or how. The June 2026 suspension of Fable 5 and Mythos 5 by government order is a real example. The models were as capable as ever. They just stopped being available.
Does this mean companies should avoid frontier models?
No. Frontier models are still very valuable, and avoiding them has its own cost. The point is to avoid depending on any single model, lab or country as permanent infrastructure. Design so a forced change is survivable.
How do you reduce AI political risk?
Set up model choice so swapping is a config change. Keep a tested fallback for production work. And track the policy side (export rules, state laws, major lawsuits) as part of the roadmap, not as background noise.
Why can't the AI lab guarantee continuity?
Some decisions aren't the lab's to make. When a government issues an export-control order, the provider must comply whatever it thinks. Anthropic did, within hours, while disagreeing in public. Continuity can depend on a regulator, not just on the vendor or on you.
Three-minute lesson on this What AI actually is Marketing course, basic. Free, no sign-up.