On June 12, 2026, the US Commerce Department issued an export-control directive that forced Anthropic to suspend Fable 5 and Mythos 5 for every customer—not a subset, everyone, effective immediately. Anthropic complied within hours. The most capable public model on Friday morning was gone by Friday evening.
Anthropic pushed back publicly. It did not change the outcome. The decision was not theirs to make.
If your workflows were pinned to that model, your vendor relationship stopped mattering the moment the directive landed. No SLA covers a national-security order. No account manager can escalate it.
Not a two-vendor problem
The obvious fix is to keep Anthropic and OpenAI on speed dial. That misses the point.
When the directive hit, Anthropic’s other models stayed online. The outage was at the model level, not the vendor level. A team that standardised on “Claude” broadly was in better shape than one that hardwired prompts, fine-tunes, and parsers around Fable 5 specifically.
Any single model can disappear without warning—through a commercial call, deprecation, a price change, or, as we saw last week, regulation. The risk is not which company you chose. It is whether one model string is load-bearing in your architecture.
What model-agnostic actually means
Saying “we could switch if we had to” is not model-agnostic. It usually means a model name is hardcoded in dozens of places and migration would take weeks.
In practice, independence looks like this:
- A routing layer between your application and the model. Agents ask for a capability—reasoning, extraction, classification—not a provider string. The platform picks the model, and that choice can change without a deploy.
- Prompts and response schemas in one place, abstracted from provider quirks. Swapping the model should not break parsing.
- Per-model fallback as default behaviour, not a last-resort catch block. When a model errors—or simply no longer exists—traffic moves elsewhere automatically.
- The same agent runnable on multiple models so you can compare cost, latency, and quality, and keep a tested alternative warm.
With those pieces in place, a Fable-5-style event is a routing change, not a production incident.
Governance, not just engineering
For regulated enterprises, uptime is only half of it. Under the EU AI Act you are accountable for the AI systems you deploy. If a model disappears and you fall back to something you never tested, you have shipped an unaudited change into a system you are legally responsible for.
Done properly, model independence means the switch is governed: routing decisions logged, model changes auditable, and you can show which model produced which output and when. Resilience and compliance are the same discipline—knowing and recording what is actually running.
Most teams treat agnosticism as a cost play and only later realise the audit trail and the fallback path should have been one system from the start.
A simple test
Ask your team: if the primary model were unavailable tomorrow at 9am, with no warning, how long until you are running normally again?
If the answer is measured in days—or “we would need to figure it out”—you do not have vendor diversity. You have a single point of failure with a logo on it.
Models will keep changing. New ones launch, old ones deprecate, and some will be removed by people who do not answer to your procurement team. The durable setup is to make the business indifferent to which model runs underneath.
That is what an orchestration and governance layer is for—the gap between deploying agents and operating them with confidence when the model landscape shifts.
Copyl routes agents across multiple models with per-agent identity, full audit trails, and EU data residency. If a model goes away, your agents keep running—and you can prove what changed. Talk to us about multi-model routing.