The new AI divide

On 9 June, Anthropic shipped Claude Fable 5 and Mythos 5, the most capable models it had ever put in front of the public. On 12 June at 5:21pm US Eastern time, a US government export control directive forced it to switch them off. Not throttle them. Not add a filter. Turn them off, for every customer on the planet, while the company worked out how to comply with an order that targeted any foreign national who might touch the model.

Three days. That was the entire commercial life of the best model an American lab had ever released.

Two weeks later it was OpenAI’s turn. The White House asked OpenAI to limit the release of GPT-5.6, and the company agreed to a staggered launch to a small set of government-approved partners. Sam Altman told staff the government would be “approving access customer by customer during this preview period,” and called the arrangement “not our preferred long-term model.” Commerce Secretary Howard Lutnick had reportedly phoned him to press for cross-agency testing before anything shipped. CNN reported the same request from the White House side.

So here is the situation as of late June. The two leading American AI labs both had their flagship launches gated by the US government inside a fortnight. One model was recalled after it was already live. The other was never allowed to ship freely in the first place.

If you run a business that builds on this stuff, that is the only headline that matters.

A model can be switched off after you build on it

For years the pitch for closed frontier models was simple. You don’t host anything, you don’t manage GPUs, you call an API and you get the smartest system on earth. The dependency felt invisible because it was just a line of code.

Fable 5 ended that illusion in 72 hours.

Anthropic had positioned the model for broad reach through AWS Bedrock, Google Vertex AI, Microsoft Foundry and its own surfaces. Teams were mid-evaluation, testing pricing and routing and deployment. Developers had wired Fable’s autonomous coding into their pipelines. Then the directive landed and all of it went dark at once, with no transition window and no grandfathering. The people who lost access weren’t doing anything wrong. They’d built on a product that a government decided, overnight, they were no longer allowed to run.

Picture an ordinary version of this. An Australian logistics company spends three months building a quoting and document-extraction system on a US frontier model. It works. It handles the invoices, drafts the customer replies, flags the odd entries for a human. The company retires the old manual process because the new one is faster and the staff trust it. Then a security researcher in another country demonstrates a narrow jailbreak, a department in Washington reads a memo, and the model the whole workflow depends on is suspended. The Toowoomba team that built nothing controversial, sold to no government, and broke no rule is back to copy-paste on a Monday morning. There is no support ticket that fixes that, because the vendor is also locked out.

That is no longer a tail risk. It happened, to a real model, this month.

This was never really about one jailbreak

The official reason was national security. The actual trigger, by Anthropic’s own account, was a “narrow, non-universal jailbreak” that amounted to asking the model to read a codebase and point out the software flaws in it. Anthropic reviewed the report the order was based on and said the capability shown was already widely available from other models, including OpenAI’s GPT-5.5. Its position was blunt: it disagreed “that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.”

Read that again. The company that built the model, the company with every commercial incentive to take safety seriously, said the government’s stated reason did not justify the action. Whether the real motive was cyber capability, leverage over the labs, or a policy being improvised in public, the effect on everyone downstream is identical. A capability you depend on can be removed for reasons you cannot see, cannot predict, and cannot argue with in time to matter.

You don’t have to pick a side in American AI politics to find that disqualifying for a system your business relies on. The point of view that matters here is the business owner’s: anything that can be switched off by a third party with no notice is not infrastructure. It’s a rental that can be repossessed.

The quiet gift to China

Here is the part Washington seems not to have thought through.

When you make American frontier models unavailable, unreliable or available only to a vetted few, you don’t make the world stop wanting capable AI. You make the world go and get it somewhere else. And the somewhere else is increasingly Chinese open-weight models you can download, run on your own hardware, and keep forever, because no government can phone a vendor and have them deleted from your servers.

That shift is already showing up. The Fable 5 shutdown was read across the industry as a moment for open-source AI, and Chinese open models from the likes of Zhipu, MiniMax and DeepSeek have been positioned squarely at the international developers locked out of the American ones. A developer in Jakarta, Lagos, São Paulo or Brisbane who can’t count on Fable or GPT being available next quarter has an obvious answer sitting on Hugging Face, and a growing number of them are downloading it.

Export controls were meant to slow China down. Applied to hosted models the world was already happily renting, they do something closer to the opposite. They tell every business outside the US that the dependable option, the one nobody can revoke, is the open model. And right now a lot of the best open models with permissive licences are Chinese. You hand the global developer base, and the standards and tooling that grow up around it, to the exact ecosystem you were trying to contain. That’s not a clever trade. It’s an own goal with a fortnight’s worth of receipts.

What this changes for the rest of us

None of this means stop using hosted frontier models. We still reach for them when the work genuinely needs the smartest system available, and for hard reasoning, long coding runs and dense analysis they’re often worth it. What’s changed is that you can no longer treat one closed model as the whole foundation and assume it’ll be there tomorrow.

The practical response is the same one we walked through in detail here, and it comes down to a few things.

Put a layer between your application and any single provider, so your code calls your own interface and that interface decides which model does the job. Keep the assets that actually make AI useful for your business under your own control: the prompts, the schemas, the retrieval indexes, the evaluation sets, the test cases and the source documents. Route work by what it needs, sending the genuinely hard problems to a frontier model and the narrow, repeatable jobs to an open or self-hosted model where control and cost matter more than a leaderboard score. And for anything sensitive or business-critical, look seriously at private and on-prem AI, where the weights live in your environment and the off switch is yours.

Most business AI work, the invoice extraction and the email triage and the quote checking, doesn’t need the model that wins the internet this week. It needs a model that’s still running next week. After June, those are no longer the same thing.

If you want to know which of your workflows would break the morning a model gets switched off, that’s exactly what the AI readiness assessment maps. Or bring us the system, the data and the deadline you’re worried about and we’ll work out where the real exposure is.