The Fable 5 saga: geopolitics has entered the chat

For nineteen days in June, the most capable AI model Anthropic had released to the public was unavailable to everyone. There was no deprecation period and no migration window. On a Friday evening it was serving enterprise customers on every major cloud; by Saturday morning it was gone. It came back on July 1. The same could now happen to any model a business depends on, so the sequence of events is worth following closely.

Three days between launch and global shutdown

Anthropic launched Fable 5 on June 9, 2026, the first public model in its new Mythos-class tier, one level above Opus. Anthropic reported the best scores yet on most of the benchmarks it published, with a clear lead in agentic coding. In Stripe’s early testing the model migrated a 50-million-line Ruby codebase in a single day, work Stripe estimated would have taken a team more than two months. It was priced at $10 per million input tokens and $50 per million output, roughly double Opus 4.8.

At 5:21pm Eastern on June 12, the US Department of Commerce sent Anthropic an export-control directive. It ordered the company to suspend access to Fable 5, and to its safeguard-free sibling Mythos 5, for any foreign national, inside or outside the United States, including Anthropic’s own non-citizen employees. Anthropic had no reliable way to check a user’s nationality in real time, so it pulled both models for every customer. Requests through AWS Bedrock, Google Cloud, Microsoft Foundry, Snowflake, Box and the Claude API all failed at once. Every other Claude model, Opus 4.8 included, kept running.

The trigger was a research report from Amazon. Its researchers had prompted Fable 5 into identifying software vulnerabilities, and in one case into writing code that showed how one could be exploited. The government’s concern: if the safeguards could be bypassed, a consumer product would work as an unrestricted cyber tool.

Anthropic disputed that reading. It said the technique was narrow, that it surfaced a handful of minor vulnerabilities that were already known, and that other public models, OpenAI’s GPT-5.5 among them, could find the same things without any export control. The company said publicly that a narrow potential jailbreak should not be grounds for recalling a model already used by hundreds of millions of people, and that applying the standard across the industry would effectively freeze new releases.

On June 26 the government approved restoring Mythos 5 for a limited set of US organisations that defend critical infrastructure, with Commerce Secretary Howard Lutnick citing safeguards for trusted partners. On June 30 the export controls were lifted entirely, and Fable 5 returned to global users on July 1 through Anthropic’s own apps and API. AWS, Google Cloud and Microsoft Foundry follow in phases.

The model came back with conditions. For the redeployment Anthropic trained a classifier aimed at the behaviour in the Amazon report, which routes flagged requests to Opus 4.8, and had the Commerce Department’s Center for AI Standards and Innovation test it. With Amazon, Microsoft, Google and other partners it is working on a shared scale for how severe a jailbreak is, scored on how much new capability it unlocks, how broadly it works, how easily it can be weaponised and how easy it is to find. Anthropic also set up a 24/7 monitoring team and a HackerOne programme for cyber jailbreaks, and committed to giving the US government earlier access to frontier models before release.

A government has now switched off a live AI model

For the first time, a government ordered a commercially deployed AI product offline overnight, and the company had no practical choice but to comply worldwide. The kill switch worked on the first attempt.

After the reversal, one analyst told Al Jazeera that the precedent is the real issue, and asked whether Washington will now need to approve every frontier release. That is not settled, but the mechanism has been used once. The executive order on advanced AI, which Anthropic says it helped shape, points toward formal government review before the most capable models launch.

The shutdown also had a competitive cost. Several tech executives and investors said that pausing a leading US model gave Chinese open-source developers time to close the gap. And it happened while Anthropic was preparing a possible IPO on a reported $965 billion valuation.

Seen from Denmark, every user was a foreign national

The directive was written about foreign nationals. For a Danish company that means all of its staff and all of its customers. There was no exemption to apply for and no restoration date, and the decision came from a US agency that European customers have no relationship with. A Danish bank or hospital running on Fable 5 would have been cut off the same way.

European discussion of AI risk has mostly been about the AI Act, which regulates what European companies do with AI. Fable 5 showed a different risk: a government outside Europe can withdraw access to the tool itself, at a few hours’ notice.

What to do if your business runs on one model

Put model access on the risk register. Finance, healthcare, SaaS and infrastructure customers lost a core service with no warning, no exception for regulated industries and no restoration date. If one model runs your customer support or your coding pipeline, a directive you cannot see coming can take it offline on a Friday evening. Plan for that as you would for a major cloud outage.

Read your contracts. Force majeure clauses written before 2026 did not anticipate an instant, government-ordered suspension across every platform. Ask your vendors what they owe you when a regulator pulls a model.

Wire in a second model. The organisations that got through the nineteen days had mapped their AI dependencies and switched to a fallback such as Opus 4.8 when the suspension hit. Those built on one model with nothing else connected were stuck. Treat model redundancy the way you already treat multi-cloud.

Expect tiered access. What came out of this is a three-level structure: public availability, a vetted trusted-partner tier (Anthropic’s Glasswing programme), and full suspension. More of the most powerful capabilities will arrive through authorisation schemes, and European companies will not be first in line.

Check deemed-export exposure. The directive targeted foreign nationals, but its effect landed on everyone, including US teams whose tools called the restricted model. If non-US staff or contractors use internal systems that send prompts or code to a frontier model, review that before the next model ships.

Fable 5 is back, with stronger guardrails than it launched with. The day it was switched off, the companies with a second model kept working and the rest waited nineteen days. That is the lesson to plan around.