The Day the Government Turned Off Your AI

Last Thursday at 5:21 PM Eastern, the US Government issued a directive ordering Anthropic to immediately suspend global access to two of their most capable models: Fable 5 and Mythos 5.

By 5:22 PM, they were gone.

No transition window. No deprecation timeline. No migration path. Just gone. For every business, every developer, every enterprise workflow on the planet that depended on them.

Within hours of the shutdown, my LinkedIn feed was full of it. Developer after developer, engineer after engineer, posting the same story in different words: they were mid-session, mid-build, mid-thought, and the model just vanished. No error message that made sense. No fallback. Just a context window that suddenly had no one on the other end.

I want to be clear about something before I go further: Anthropic is a world-class team doing genuinely important work. They disagreed with the directive. They said so publicly. I respect that.

But the directive happened. And it will happen again.

This is the new shape of conflict

We have spent the last two years watching governments weaponise economic infrastructure in ways we have not seen since the Cold War. Tariffs. Trade restrictions. Supply chain decoupling. Semiconductor export controls. The message has been consistent: strategic technology is no longer a neutral global resource. It is a lever.

AI capabilities are now that lever.

What happened to Fable 5 and Mythos 5 is not an isolated incident of a government exercising caution over a jailbreak concern. It is the first visible use of AI model access as an instrument of national policy. The justification this time was a vulnerability. Next time it may be geopolitical friction. A trade dispute. A diplomatic falling out. A sanctions regime that catches AI providers in its net the same way semiconductor manufacturers have been caught.

We are not speculating. The architecture for this kind of control already exists. The Bureau of Industry and Security has export control authority over software. The Entity List already restricts what American technology companies can provide to certain foreign nationals and organisations. The EO on AI from 2023 established reporting requirements that can be tightened. The tools to restrict AI model access globally, selectively, and overnight are already in place. Thursday was simply the first time we saw them used.

“The question is not whether this will happen again. The question is which jurisdictions, which models, and which enterprises will be caught in the middle when it does.”

The parallel to trade wars is not abstract

When the first round of tariffs landed in 2018, most manufacturers had not seriously stress-tested their supply chains for a scenario where a single government decision could make their inputs disappear overnight. They had optimised for efficiency. They had not designed for resilience. The ones who recovered fastest were the ones with diverse supplier relationships and the ability to switch.

Enterprise AI is in exactly the same position today that global manufacturing was in 2017.

The organisations that built directly on one provider's API, that trained their workflows on one model's output patterns, that made one company's capabilities the foundation of their operations, are the organisations that will feel the next directive most acutely. They have optimised for the capability of a single model. They have not designed for the day that model becomes unavailable.

This is not hypothetical risk management. This is the lesson that Thursday just taught, and most of the enterprises it affected have not yet finished absorbing it.

I am not suggesting that Thursday's directive was an act of aggression. It was not. But the capability it demonstrated, the ability to remove access to AI infrastructure for every organisation on the planet simultaneously, is one that other actors are watching. And the fragility it exposed in enterprises that have not diversified their AI dependencies is one that should concern every board, every CTO, and every CISO who read the news that night.

Why we built OBEL

There were two primary reasons we built OBEL.

The first was to provide a platform that is model and provider agnostic, while delivering a consistent security layer, observability, governance, and an immutable audit trail. We believe it is essential for enterprises to have the ability to utilise the best models available behind a controlled proxy. That is not a preference. It is a requirement for any organisation that takes its obligations seriously.

The second was ARGUS-i. Our runtime engine powers everything. It classifies, scrubs, governs, and audits every interaction regardless of which model sits underneath it. When you swap a model, ARGUS-i does not care. The security layer does not change. The audit trail does not change. The compliance posture does not change.

Powered by ARGUS-i, OBEL is exactly that: a controlled, governed proxy in front of any model, from any provider, including open-weight models you can run on infrastructure you own.

“You don't own the model. You never did.”

- Denis Bouton, Managing Partner, ninthLABS Ventures

When you integrate directly with an LLM provider's API, you are building your business on infrastructure you don't control, in a jurisdiction you may not operate in, subject to regulatory decisions by governments that have no obligation to give you notice.

That is not a criticism of any provider. It is simply the architecture of the current AI landscape.

What happened when Fable 5 went dark

The open-source gap has closed

For three years, the enterprise AI conversation was dominated by closed, proprietary models from a handful of US-based labs. The performance gap justified the dependency.

That gap is now closed.

Llama 4 Maverick. DeepSeek R1. Mistral Large. These are not compromise alternatives. They are genuinely excellent models you can deploy on infrastructure you own, in jurisdictions you choose, with no vendor who can pull access on 60 seconds notice.

Models available in OBEL today

Every model below is available inside OBEL right now, governed through ARGUS-i, with no workflow changes required when you switch between them.

• Anthropic: Claude Sonnet 4.6, Claude Haiku 4.5, Claude Opus 4.8. Still available. Fable 5 and Mythos 5 remain suspended.
• Google: Gemini 3.5 Flash, Gemini 3.1 Pro Preview, Gemini 2.5 Pro, Gemini 2.5 Flash. Different regulatory jurisdiction.
• OpenAI: GPT-5.5, GPT-5.4, GPT-5.4 Mini, GPT-5.4 Nano. Reasoning now integrated into GPT-5.4 Thinking mode.
• Meta Llama 4: Llama 4 Maverick, Llama 4 Scout. Open weights, multimodal, self-hostable on your own infrastructure.
• xAI: Grok 4.3, Grok 4.20, Grok 4.1 Fast. Different risk profile, generous free API credits.
• Mistral: Mistral Large 3, Mistral Medium 3.5, Ministral 8B. European jurisdiction, open-weight options available.
• DeepSeek: DeepSeek V4 Flash, DeepSeek V4 Pro. Open weights, 98% prompt caching discount, self-hostable.
• Groq (free tier): Llama 4 Maverick, Llama 4 Scout, Kimi K2, Qwen3 32B, GPT OSS 120B. No API key required.

OBEL customers point their agent or workflow at their preferred model and ARGUS-i routes intelligently. No workflow changes. No re-integration. Just resilience.

The question every enterprise should be asking

This is the moment to re-examine your AI stack. Not because something catastrophic happened to you last Thursday, but because something almost did. The question is simple: what is your plan if it happens to you next time?

OBEL exists to be that plan. The governance layer, the security proxy, the multi-model routing, the audit trail, the human-in-the-loop gates, the workflow orchestration. All of it runs on any model, from any provider, or on your own infrastructure.

“The model is not the moat. The governance is the moat.”

If you are an enterprise leader reading this, I want to have that conversation with you. Not to sell you something. To show you what model-sovereign AI infrastructure looks like in practice.