On Friday, June 12, a frontier model went dark. A customer didn’t shut it off. An engineering team didn’t decide to switch providers. A government did.

The US government issued an export-control directive on Claude Fable 5 and Mythos 5, suspending access for any foreign national, inside or outside the country. Anthropic couldn’t verify nationality in real time on every request, so it did the one thing that complied: it disabled both models for everyone. Paying US customers included.

That’s the thesis, and it comes before any detail: if your operation runs on a capability you don’t control, you don’t own a tool. You own a continuity risk that simply hasn’t shown up yet. IQ Source’s concrete piece for companies already carrying that exposure is Socio Tecnológico. The rest of this post is about why “just go open source” answers the wrong question.

What got switched off, and what that proves

The facts matter because they define the kind of risk. This wasn’t a server outage or a rate limit. It was an external decision, outside the contract between Anthropic and its customers, that made a capability unreachable overnight.

Aaron Levie, CEO of Box, framed it well when he wrote on X that the layer able to route to the best model for each job is about to gain a lot of value, for three reasons: cost optimization, capability maximization, and risk mitigation. The first two we already knew. The third stopped being theoretical on Friday. As he put it, we may be heading toward a regulatory environment where governments restrict models at different times, so you’ll want the flexibility to move workloads across providers as a form of risk mitigation.

Yang Li, who builds sovereign AI infrastructure, said it more bluntly on LinkedIn: the industry won’t stop talking about how smart the model is, and almost nobody talks about who can switch it off. The most important question stopped being “how capable is it?” and became “who controls the switch?”

You don’t have to buy the national-sovereignty argument for the operational conclusion to apply to you. A fifty-person company in San José that classifies tickets, drafts proposals, and runs three internal agents on a single model faces the exact same structural problem as a nation-state: it built on a capability whose switch sits in someone else’s hand.

The mistake isn’t using Claude. It’s having no fallback.

The loudest reaction this week was “this proves you have to go open source.” It’s the wrong answer to the right problem.

Moving everything to an open model doesn’t buy you independence. It buys you a different provider of the same problem: now you depend on whoever hosts it, on that service’s limits, on its availability. You changed who owns the switch, not how many switches there are. The fragility was never that the model belonged to Anthropic. It was that your operation had a single point of failure and nothing ready to start in its place.

What changes the math isn’t the model’s license. It’s where the logic that decides which model handles each task lives.

I made this point from a different angle in the post on the cheap model as bait: the difference between choosing your next provider and having it chosen for you is, technically, about fifty lines of code. If the routing logic sits in your provider’s console, the provider controls what scales, what gets served, and what happens when something breaks. If it sits in your code, a function that evaluates the task and decides which endpoint to call, then switching from Claude to another model is a config change, not a three-week project.

The proof that the multi-model approach already works in production landed the same week. OpenRouter shipped its Fusion API and published the numbers: a panel of three budget models, fused together, beat solo GPT-5.5 and solo Opus 4.8 outright, and landed within one percent of Fable 5 at roughly half the price. I’m not citing it as a product recommendation. I’m citing it as evidence that orchestrating several models is no longer a lab experiment. The ability to move work across models is mature. What most companies lack isn’t the technology. It’s the design.

The abstraction isn’t free

This is where most LinkedIn posts stop, right before the uncomfortable part. If this were as easy as “add an abstraction layer,” everyone would already have one. They don’t, for real reasons.

First, caching. A lot of the savings in tools like Claude Code come from context being cached across calls, and that cache breaks the moment you switch models. Moving a workload to a backup provider in an emergency isn’t free: you pay to rebuild the context, and sometimes you eat higher latency while you do. Keeping the door open costs something, even when you don’t walk through it.

Second, governance. Leon Gordon put it as a question with no clean answer: when a vendor’s compliance posture can trigger a global takedown with no notice, how does that sit inside an SLA or a procurement risk framework? Running live integrations with several providers means several SDKs, several response formats, several pricing models, several sets of limits. The goal isn’t to use every model on the market every day. It’s to be able to switch in hours when you need to. There’s a difference between designed redundancy and three unmaintained integrations nobody owns.

Third, and this is the one that weighs most: someone has to own that layer. Design it, test the fallback for each workflow, patch behavior when a model changes, decide which workloads deserve redundancy and which don’t. Most mid-market B2B companies in LatAm don’t have that role in-house. They have engineers who can wire up a model and make it work. They don’t have the person who sits on top making sure the operation survives the day someone else moves the switch.

The cost of that fragility is already being measured on the trust side. David Eberle, CEO of Typewise, cited a number on LinkedIn that reads differently after this week: team trust in AI drops from 82% during the build phase to 58% at release. An external shutdown with no warning is exactly the kind of event that widens that gap. It doesn’t punish the model’s capability. It punishes the assumption that it will be there tomorrow.

What IQ Source does about it

Socio Tecnológico exists to be that owner. The client company doesn’t outsource wiring up a model to us; that part became accessible. They outsource the role that designs the routing layer, maintains a fallback for each critical workflow, absorbs the change when a provider fails or raises its price, and keeps that knowledge inside a trained team instead of one engineer who might quit. It’s the role a good CTO would fill, the one most mid-market companies can’t afford full time.

The prior discovery is handled by AI Maestro, our two-month program that produces the Process Reality Map and the AI Opportunity Score. That’s where a decision gets made that matters as much as the redundancy itself: which workflows deserve a fallback and which don’t. Not everything running on a model is critical. Putting redundancy on a task that tolerates a day of downtime is paying for complexity you don’t need. The discipline is knowing where continuity is negotiable and where it isn’t, and you make that call before you build, not during the outage. We worked the same logic from the cost angle in the post on the moat that moved to the workflow.

Ask your team one concrete question before the week closes. If the model your most critical workflow runs on disappeared tomorrow because of a decision you didn’t make, how many hours until you’d be running on another? If the answer is measured in weeks, or if nobody knows it, you don’t have an AI tool. You have a dependency with its switch in someone else’s hand. And this week made it clear that hand moves.

Design the fallback for your critical AI workflows