Thank you for opening up the latest article in 'The Intelligent Commit' newsletter, my thoughts, research and analysis on Artificial Intelligence and how AI is reshaping the software development lifecycle. I hope you find it insightful, subscribe, and engage.

I spent a good chunk of my career leading Site Reliability Engineering teams, and SRE gives you a particularly pessimistic way of looking at the world. Everything fails eventually, so you design for resilience rather than hope failure stays away. We ran disaster recovery drills, we broke services deliberately, and we migrated services east/west and back a week later, all in office hours, so failures wouldn't surprise us at 3am. We tested the 'backup generators', metaphorically speaking.

We routinely tested our architectures by asking ourselves what happens if a critical dependency fails? Yet in all those years, I can't remember anyone ever asking what happens if a dependency is withdrawn.

Last month, that stopped being a hypothetical question.

On 12th June, three days after launching Claude Fable 5, Anthropic received a US government export-control directive and suspended access for every customer, everywhere. Those controls were eventually lifted and global access returned on 1st July. While we all rejoiced in our offices and spent the weekend completing our hobbyist projects with the extra usage credits, let's not forget that the Fable blackout lasted 18-19 days.

This wasn't a traditional downtime. There was no SLA to invoke, no service credits, no status page promising a fix. A regulator decided, rightly or wrongly, and almost as soon as it arrived the most capable model on the market disappeared.

Keep Calm. On paper, the industry passed the test.

Let me make the case for calm first, because it's a very solid case. Especially when Fable was too new to have made its way into widespread production use.

VentureBeat Pulse Research surveyed 145 enterprises over the blackout period, and found two-thirds already had a multi-model hedge in place before the order came down. Routing requests across multiple models, sized to task criticality and cost, is becoming a standard pattern. Of surveyed enterprises, 51% blend closed frontier models with open-weight models on their own infrastructure (more on this later), and another 16% are moving core workflows off closed APIs entirely.

And even if Fable had not been a newcomer, which dampened impact significantly, the market filled the gap quickly. Sakana AI pitched its Fugu system as "frontier capability without export-control risk" within days, and OpenAI launched GPT-5.6.

Eighteen days, hedges held, nobody died. Case closed, nothing to see here, right?

Not quite, because the hedge may have covered the wrong asset.

If we dig a little deeper into that same VentureBeat survey the picture changes. It reveals only 1 in 10 enterprises has automated monitoring that would catch their AI models drifting, misbehaving, or failing in production, and roughly a quarter would learn of a failure from their end users, or wouldn't detect it at all.

So even though two-thirds hedged the API, almost none hedged the business workflow. That gap highlights that your real dependency on a model vendor is the behaviour you've shaped your business process around, not the endpoint you call i.e. the prompts tuned to one model's quirks, the agent scaffolds, the evaluation suites, the skills and guardrails calibrated against specific outputs. And that's the key lesson we should take away from those 18 days without access to Fable.

If you're reading this in the UK and filing it under "American problem", Westminster would like a word.

While Fable was dark, MPs here in the UK were debating an off-switch of their own. An amendment to the Cyber Security and Resilience Bill, tabled by Leeds MP Alex Sobel and backed by MPs across party lines, would grant the Secretary of State emergency powers to shut down data centres and AI systems that pose a serious risk to public safety, essential services, or national security. The threshold is deliberately high, and the government hasn't endorsed it. But the direction is clear as we witness the power to withdraw an AI system being written into legislation on both sides of the Atlantic.

I'm not arguing the amendment is wrong; there are serious safety arguments for it. I'm arguing that your threat model now has to include the off switch, whoever's hand is on it. In June it was in Washington, but the next one may be rather closer to home.

A backup model only counts if it's available where your data has to live.

There's a sovereignty wrinkle that failover plans need to consider too. When AWS went GA with Bedrock AgentCore, its agent hosting platform, it launched in nine regions, and London, UK wasn't one of them. We had to wait until January for the core capabilities, and until March for some of the rest. So if your data residency obligation requires that you keep workloads in the UK, your menu of fallback models and platforms may be shorter. A failover plan built on what exists in us-east-1 isn't a guaranteed plan for a UK-regulated business, as I found regularly when architecting AWS and GCP cloud services via a patchwork of available options across London, Frankfurt, and Belgium.

And hosting your own is a hedge most SMEs can't afford.

I promised to return to an earlier point. The survey's most popular hedge, blending in open-weight models on your own infrastructure, quietly assumes you have infrastructure. Academic work on on-premise LLM deployment finds the economics only break even against API pricing at sustained, very high utilisation, once you account for GPU capital costs and the expertise to run them. If you're an enterprise with a platform team and a data centre contract, fine. If you're one of the thousands of UK SMEs building on AI with neither, sovereignty-through-ownership is simply out of reach, requiring a proportionate and pragmatic response.

That doesn't leave smaller organisations helpless, but it definitely means their hedge looks a lot different, built from a second API provider, evaluation suites that aren't welded to one model, and failover that has been exercised at least once. Which brings me back to the discipline I started this article off with.

An untested backup is not a backup.

Have you run failover drills on your AI models? Will you? The recent Fable blackout, like the loss of access to AI code assistants I wrote about in my previous article, didn't create a new weakness. But hopefully it has shone a light on a blindspot, and highlighted an opportunity for some to strengthen their operational resiliency for the next, inevitable, 'off switch' moment here in the UK or elsewhere.

What this means for engineering leaders.

I suggest four places to start, in order.

  1. Tier your AI-enhanced business workflows by criticality, the way you would any other dependency, and ask yourselves whether losing one for a day would be an inconvenience or a P&L event.
  2. Build the monitoring before the multi-model strategy. You cannot fail over from a system you can't see drifting or failing, and the VentureBeat survey data says this is where 90% of organisations are currently exposed.
  3. For your top tier only, implement real portability, keep evaluation suites model-agnostic, and run an actual failover exercise, such as a game day where the primary model is unavailable and the system or team switches. While full feature parity may be unaffordable or unachievable, exercising your failover plan for the business workflows that matter most is not.
  4. Map your fallback's geography and read the terms you've already signed. Does your second model exist in a UK or EU region? Where does the data go, who can revoke access, and under which jurisdiction?

Those 18 days turned out to be survivable, and I repeat that Fable was too new to matter to most, but that isn't the point of this article. But survivability is not the same as being resilient, and nothing about those 18 days was guaranteed. The next directive, from whichever government, could last longer.

Thank you for reading my latest article. Please share your experiences and engage with the discussion. I am not always right, but I try to be insightful and thought-provoking.

  • Did the June blackout impact your organisation? Beyond the news cycles, how quickly would you have known?
  • Would a UK kill switch change your risk register, or does your threat model already include your own government?
  • Have you ever exercised an AI model failover for real, or does the second vendor live only on a PowerPoint slide?
  • If you're an SME without your own infrastructure, what does a proportionate hedge actually look like for you?

Please reach out if you'd like to continue the discussion.

Sources

  1. Anthropic, "Statement on the US government directive to suspend access to Fable 5 and Mythos 5" — https://www.anthropic.com/news/fable-mythos-access
  2. Forbes, "Anthropic Disabled Fable 5 And Mythos 5 After A US Export-Control Order. Here's What Happened" — https://www.forbes.com/sites/anishasircar/2026/06/16/anthropic-disabled-fable-5-and-mythos-5-after-a-us-export-control-order-heres-what-happened/
  3. CNBC, "Anthropic says Trump admin has lifted export controls on Claude Fable 5 and Mythos 5" — https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html
  4. Anthropic, "Redeploying Claude Fable 5" — https://www.anthropic.com/news/redeploying-fable-5
  5. VentureBeat Pulse Research, "Enterprises lost Claude Fable 5 for a few weeks. New data shows two-thirds had already built their hedge" — https://venturebeat.com/orchestration/enterprises-lost-claude-fable-5-for-a-few-weeks-new-data-shows-two-thirds-had-already-built-their-hedge
  6. VentureBeat, "No Claude Fable 5? No problem: Sakana achieves frontier performance with new Fugu multi-model system" — https://venturebeat.com/orchestration/no-claude-fable-5-no-problem-sakana-achieves-frontier-performance-with-new-fugu-multi-model-auto-synthesis-system
  7. Computer Weekly, "MPs propose 'kill switch' to shut down rogue AI systems" — https://www.computerweekly.com/news/366643176/MPs-propose-kill-switch-to-shut-down-rogue-AI-systems
  8. AWS, "Amazon Bedrock AgentCore release notes" (GA October 2025 in nine regions; Europe (London) added January 2026) — https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/release-notes.html
  9. "A Cost-Benefit Analysis of On-Premise Large Language Model Deployment: Breaking Even with Commercial LLM Services" — https://arxiv.org/html/2509.18101v1