The Fable 5 ban lesson is to diversify

Last month, Washington ordered Anthropic to block every foreign national from its most capable public model, Claude Fable 5. The company could not separate foreign users from American ones in real time, so it disabled the model for everyone. Australian businesses that had started building on Fable 5 woke up to an error message. It is tempting to read that as an alarming new precedent. It is not.

The kill switch is not new, and the rest of the world is already acting on it. Washington sanctioned the International Criminal Court's chief prosecutor last year and Microsoft cut off his email, a service his office used every day. France has since ordered its ministries to shift 2.5 million government computers off Microsoft software and onto open-source alternatives. Australia has now had its own version of the lesson. The genuinely new development is quieter, and more useful to a business owner than another round of alarm.

While everyone watched Fable 5, the open models caught up. Stanford's AI Index measured the gap between the best open and the best closed model on the main public leaderboard at 8 per cent at the start of 2024. A year later it was 1.7 per cent. Models from z.ai GLM5.2 and OpenAI's open release GPT-OSS now match the paid flagships on many standard benchmarks. For the everyday work most businesses run, the gap has closed.

They are also far cheaper. Fable 5 costs roughly double Opus 4.8 for the same volume of text, and the open models sit well below both. A leading open model runs at close to a tenth of the price of a flagship frontier tier. For a small business running AI across email, meeting transcription and customer service, that is the difference between a tool it has to ration and one it can run everywhere.

Put those two facts together and the lesson is not that every business needs a sovereign model of its own, or even that it should fear the off switch. It is simpler. No company, large or small, should build itself around a single model. The clinic that wired every workflow into one frontier system did not have a sovereignty problem; it had a concentration problem.

You do not need a foreign government to pull a switch for single-model dependence to hurt. The same exposure shows up in far more ordinary ways. A provider can raise its prices, and frontier pricing has jumped sharply before. It can retire the exact model your workflow was tuned to, leaving you to rebuild on short notice. It can have an outage on a Tuesday morning, or throttle you with a rate limit in your busiest week. A business that can fall back on a second model rides all of that out. A business wired to one simply stops.

Something can finally be done about it. Until recently, diversifying meant opening a second account with another provider in the same country, which is barely diversification when both answer to the same government. The capable open models change that, because being open-weight they can run on a different cloud, on local servers or on hardware inside the building. Switching is not free, and a different model behaves differently enough that you test it before you trust it, but that is a planning task rather than a barrier. Real diversification is no longer only about which model a business uses, but about where the model runs and who can reach the switch.

For many of the businesses I work with, this meets an older concern. They have never been comfortable sending private client records to a service they cannot see inside, unsure whether that data ends up training someone else's model. A capable model running on infrastructure they control settles the question. It cannot be switched off from 14,000 kilometres away, and the data never leaves the building.

None of this means a small business should stand up its own model server this week. Diversification is a spectrum, not a weekend rebuild, and the first step costs almost nothing. Stop hard-wiring your workflows to a single provider, keep one alternative you have actually tested rather than one you assume will work, and treat the model as a component you can swap rather than the foundation you pour everything onto. From there a business can move at its own pace, adding a local or Australian-hosted option as the work and the budget justify it. The point is to keep the exit open.

The kill switch was always there. What has changed is the alternative, capable open models that are cheap enough and good enough to run yourself. Anthropic expects the restrictions on Fable 5 to be lifted before long, and the model will probably return. But the past week is the signal to take diversification seriously, while the choice is still a calm one rather than a forced one. When governments treat AI as a national security matter, reliance on any single model is a political variable, not a fixed asset.

A model you rent can be switched off from afar. A model running on your own hardware cannot.