On 12 June, four days after launch, Anthropic switched off Claude Fable 5 and the Mythos 5 model beneath it for every customer on earth. It had little choice. A US Commerce Department export-control directive ordered the company to deny the models to any foreign national, inside the United States or out, including its own non-citizen staff, and the cleanest way to comply was to pull them for everybody. The stated reason was a jailbreak that turned Fable 5 into a capable cybersecurity weapon. The most powerful model in the world had been public for ninety-six hours.

For a founder in Singapore, Jakarta, Kuala Lumpur or Manila, the useful lesson is not the one the region will reach for first. The reflex will be the sovereignty essay: governments fretting about being cut off, fresh calls for a Southeast Asian foundation model, a panel at the next conference. Leave that to the policy desk. The smaller read is about supply. You do not own a capability you reach through one foreign vendor that can be ordered to cut you off on a Friday; you rent it, and the landlord answers to a government that is not yours.

You do not own a capability you reach through one foreign vendor that can be ordered to cut you off on a Friday; you rent it, and the landlord answers to a government that is not yours.

This is the first time Washington has applied an export control to a model rather than to the chips that run it. Opus 4.8 and the lesser Claude models kept working, and so did the other frontier labs; the blast radius was one tier of one vendor. The precedent is the asset here. The outage itself was almost a footnote.1 A frontier capability can now disappear for non-Americans with no notice and no appeal, for a reason the user had no part in creating and cannot fix.

Geography is where this gets uncomfortable. The directive reached inside the United States, to H1-B engineers and to Anthropic’s own foreign staff, which means hosting your workload in a Singapore data centre buys you nothing. The control travels with the passport, not the server. A regulated-jurisdiction host, the thing the region has spent two years selling as insurance, is irrelevant to an order written this way.

The control travels with the passport, not the server.

So the move is the one Charlie already runs in the durian trade, where you never source from one farm in one season and pray the weather holds. Pick a primary model for the work that pays the bills, the drafting, the summarising, the customer-facing layer, then keep a tested fallback on a different jurisdiction’s model: a Chinese open-weight release such as GLM or Qwen, or a local-host option that no foreign directive can switch off. The testing is the point; owning was never on offer. A fallback you have never run is a line on a slide, not a supplier.

The skeptic has a case, and it is a fair one. This hit a single model for a single narrow reason; five other top-tier models still serve the region perfectly well; and a durian-and-oysters operator wiring up a Chinese-model backup against the small chance Washington acts again is insuring a house that will probably never burn. Mostly right. So the honest version is modest: you learned this week that the cut-off can happen, cheaply, with no warning, and the insurance against it costs you one quiet afternoon of testing a substitute you hope never to need.

The model was the best in the world for four days. What it proved will outlast the outage. The operators who come through this unbothered are the ones who already knew, before 12 June, which single supplier their business could not survive losing, and had quietly run the second one once, just to be sure it worked.

Footnotes

  1. Within days a prediction market opened on when the ban would be lifted. The order named no date; by mid-June, neither had Anthropic. ↩