By Damian Mathews & The Last Mile Team
Claude Fable 5 launched on June 9. I started using it that day, and could feel the difference right away. It was, by a clear margin, the best model I had ever used.
I had it for three days.
On June 12, a US export-control directive forced Anthropic to pull Fable 5 and its sibling Mythos 5 offline, worldwide. Off for everyone, including paying subscribers who had the best model of their lives on Thursday and an error message by Friday.
Four days later it topped the public AI benchmarks as the highest-scoring model anyone had tested. The best AI model on the planet, sitting at number one, and completely impossible to use.
Anthropic didn’t choose this. The directive told it to suspend access for any foreign national, which in practice meant everyone. The company said it disagreed and was working toward a fix, but gave no timeline, because it doesn’t control one. Its other models stayed live. Fable 5 and Mythos 5 went dark and remain dark.
For any company that had built Fable 5 into a live workflow, the disruption arrived with no warning. Nothing had gone wrong in their own systems, their cloud, or their billing. A government action the vendor itself contested and lost had reached in and switched off a capability overnight.
In $1.22 on the Dollar, we argued the AI model layer has become a volatile dependency any single-provider contact center inherits in full. Last week that risk was commercial: pricing, sunsets, roadmaps bending toward Wall Street. This week added one almost nobody had on their board: regulatory and geopolitical. A model can now vanish for reasons unrelated to your uptime or spend.
That’s a new kind of exposure most AI governance frameworks don’t account for. Teams evaluate models for capability, cost, and safety. Very few ask what the recovery posture looks like if the model disappears, or whether access might suddenly vary by an employee’s nationality, as this directive just produced. For any enterprise running offshore teams or BPO partners, that second point deserves more than a footnote. If your agents in Manila or Bogotá lost access to a capability your US team still had, the containment gap shows up in your numbers before anyone has finished reading the directive.
And here’s the part that should concern you even if a government directive feels like a remote scenario: the mundane version of this happens all the time. A CCaaS platform quietly deprecates a native AI feature with a quarter’s notice. A model update breaks prompt behavior you’d spent months tuning. A vendor roadmap pivots away from a capability you’d built a workflow around. The Fable 5 story is dramatic. The underlying condition, capabilities entangled with dependencies you don’t control, is ordinary. Most contact centers that have deployed gen AI in the last 18 months are already living it.
Think about what that means for a live contact center. Containment flows stop working, routing reverts to static logic, agent assist goes dark mid-shift. Every one hits a KPI your team owns: containment, transfer rate, AHT, CSAT. And the accountability sits with a government directive or a vendor deprecation notice, which means there’s no escalation path, no credits, and no timeline. When the capability that fails isn’t yours to fix, the SLA gap becomes yours to absorb.
The real exposure here is bigger than the model failing. Everything built on top of it failed with it. That’s the dependency problem: when your containment logic, your routing intelligence, and your agent playbooks are native to a single model or platform, they don’t survive a change you didn’t choose. And the changes you don’t choose are the ones that arrive without warning.
Most teams don’t solve this upfront. The platform makes it easy not to: building above the model layer takes more architectural discipline at the start, while native AI bundles are faster to deploy and easier to demo at go-live. So that’s what gets built. And it works, until it doesn’t.
The way we work comes down to this: prove the capability first, then keep it portable. That means building containment flows and agent logic that sit above the model layer, not inside it, so that swapping the underlying model is a configuration change, not a rebuild. It also means owning the orchestration ourselves, so when something changes beneath us, we’re not waiting on a platform vendor’s roadmap to respond. The KPI ownership doesn’t move. The capability does.
None of this is a shot at Anthropic, which makes brilliant models. That’s the unsettling part. The lab at the top of the index still couldn’t keep its best model online. If it can happen to the best model in the world, it can happen to whatever is in your stack.
I’ll get over Fable 5 eventually. 😔 Or maybe it’ll come back.
If your AI-enabled CX capability disappeared tomorrow morning with no warning, how long until your agents were back up?
Most teams we talk to don’t have a clean answer to that. The answer lives inside a platform they don’t fully control, wired to a model they didn’t architect around, built by an integrator who handed it off at go-live and moved on. That’s the conversation worth having before something forces it. CX Foundry is where we start: we analyze your stack and customer interactions, prove what your contact center can actually do, and make sure the capability stays yours when things change beneath you.
The companies that weathered this week with no disruption had already made the architectural decision that everyone else is now being pushed toward. The only question is whether you make it on your own timeline or someone else’s.
— Damian
Here’s what went down this week.
Bleeding Edge
Early signals you should keep on your radar.
SpaceX agreed to buy Cursor for $60 billion in all-stock, days after its record IPO. Cursor’s parent, Anysphere, runs near $2.6 billion in annualized revenue, and the two plan a shared model spanning Cursor and Grok. Folding a frontier coding lab into Musk’s rocket empire is a big swing, and the combined model could reshape his AI stack.
Sharon AI signed a six-year NVIDIA pact to stand up 40,000 Grace Blackwell GPUs in Australia. The roughly $4.88 billion deal adds 72 megawatts of capacity under NVIDIA’s DSX “AI factory” design. Regional compute hubs like this suggest the GPU buildout is spreading well beyond the usual US and Gulf megaprojects.
Leading Edge
Proven moves you can copy today.
TCS partnered with Anthropic to put Claude in front of 50,000 of its own staff and build products for regulated industries. The firm will package Claude into offerings like insurance claims processing and bank lending advisory across 56 countries. Treating itself as “customer zero” is shrewd, since clients tend to trust an integrator that has already run the playbook.
KPMG expanded its Microsoft alliance to deploy Agent 365 and Copilot for managing AI agents across client organizations. The rollout folds Microsoft’s agent tooling into KPMG’s Trusted AI framework for monitoring and securing agents at scale. As companies move from pilots to fleets of agents, governance is becoming the product, and that shift looks hard to reverse.
Off the Ledge
Hype and headaches we’re steering clear of.
The US government forced Anthropic to disable Claude Fable 5 and Mythos 5 just three days after launch. An export-control directive barred foreign nationals from the models, so Anthropic shut them off for everyone to comply. Other Claude models stayed online. The lesson isn’t about Anthropic — it’s about any single-model dependency in a live CX stack. If the best-scoring model in the world can go dark overnight, architectural resilience isn’t optional.
A German court held Google liable for false AI Overviews that tied two publishers to scams they had nothing to do with. The Munich court called the summaries Google’s own content, stripping the safe-harbor protection search results usually enjoy. Google is appealing, but if “the AI said it” stops working as a defense, every vendor shipping generated answers should take note.
