AI & Extreme Ownership
Rent the Frontier, Own the Stack
On June 9th, Anthropic made generally available the most capable model ever built. Three days later, it was taken back.
Fable 5 is the safety-guarded version of Mythos: the same underlying model, with guardrails that quietly route the riskiest queries to a weaker fallback.
It is, so far, the most powerful model I’ve ever used. And also, very different.
Different in the what, the kind of task worth giving it to extract maximum value, and also different in the how: a clear goal, hard constraints, deep context, then room to run.
Early adopters gave it work that used to take teams. Then it vanished.
The order was to suspend access for all foreign nationals on national-security grounds, apparently over a narrow jailbreak the company believes was a misunderstanding. Anthropic couldn’t cleanly separate who was who, so it shut the models off for everyone.
In an instant, everyone learned that the frontier synthetic intelligence we built workflows, teams, and systems around was never ours at all.
The company that built it, a government, or a supranational authority can take it away in an instant. No transition period, no warning.
The reasons can be varied, and you may never learn which one applied: geopolitics, military advantage, a commercial edge, the raw economics of who gets the compute, or a feud between a lab and the state.
Jocko Willink built a leadership doctrine on one sentence: leaders must own everything in their world, there is no one else to blame. It’s a useful test to point at your AI stack.
Jensen Huang has been making a version of this at the level of nations for a couple of years.
Every country needs to own the production of their own intelligence. You cannot allow that to be done by other people.
Jensen Huang
Swap “country” for “company” and you have the brief for the rest of this piece. You probably can’t own the frontier. You can own almost everything around it, so that the frontier becomes a component you rent rather than a foundation you stand on.
Hold the frontier, and use it without mercy
Owning your stack starts with using the best models hard while you still have them.
When a Fable-class model is in your hands, give it the work you’ve been avoiding: the codebase migration, the customer-behavior analysis nobody has time to run, the GTM teardown, the legacy system no one understands.
The ceiling is higher than people assume. Stripe reported that Fable 5 ran a codebase-wide migration on a 50-million-line Ruby codebase in a day, work that would otherwise have taken a team more than two months by hand.
The second move is less obvious but more durable. Have the frontier model audit and rebuild the harness around your cheaper models, so the rest of your stack inherits some of that intelligence after the frontier is gone.
I keep meeting teams that underinvest here. The code that decides what context a model sees, what tools it can call, and how its work gets checked turns out to matter enormously. The same model can show up to 6x performance variation from the design of that scaffolding alone. A well-built setup on a mid-tier model can beat a naive setup on a better one:
Meta and Harvard’s Confucius Code Agent posted 52.7% on SWE-bench Pro running Claude Sonnet 4.5 rather than the top model.
Treat your harness as the asset that outlives any single model.
But...
What do you actually own when your smartest tech lives on someone else’s servers?
The resilience stack
You answer it with architecture. The goal is a middleware layer and a set of business logic that let you choose, route, and fall back across the entire intelligence market, so that any one provider, model, or jurisdiction can go dark while your operation keeps running.
The gateway. A model gateway is a single proxy that gives your application one API across many providers, with routing, automatic fallback, caching, and cost controls built in.
Instead of wiring your code to each lab’s standard, you wire it to the gateway once.
Self-host with LiteLLM, which fronts more than 100 providers behind an OpenAI-compatible API; reach for a hosted marketplace like OpenRouter and its 300-plus models; add governance and observability with Portkey; or use the hyperscaler versions you may already be paying for, AWS Bedrock and Microsoft Foundry.
When a model disappears (or gets too expensive, or underperforms), you change a line of configuration instead of rebuilding for a quarter.
Open weights and owned inference. The gateway only matters if it has somewhere safe to route to. Open-weight model families (Llama, Qwen, DeepSeek, now GLM) close much of the gap with closed frontier models on the shorter, well-scoped tasks that make up most production needs.
You can rent fast inference for them from Groq or Cerebras, or stand up your own.
Either way you hold a capability floor that no vendor and no government can revoke, because you are running the weights.
Portable frameworks. Your orchestration logic, the part that defines how agents plan, call tools, and hand off, can be independent of one lab’s SDK.
Framework-level tools like LangGraph and CrewAI run across providers by design. The lab-native kits (OpenAI’s Agents SDK, Anthropic’s, Google’s ADK) are excellent and optimize for one house. Use a portable option for the spine of your system and reach for a native one only where a specific capability earns the dependency.
The portable tooling. Around every serious agent sits a ring of supporting systems that compose its harness: memory, retrieval, tools, state management, error recovery, security, observability and evaluation.
You may choose ones that treat the model as a swappable input: Mem0, Zep, or Letta for memory; LlamaIndex for retrieval; Langfuse or Helicone for observability; Promptfoo for evaluation. If your evaluation suite only understands one provider’s outputs, you can’t actually compare alternatives, which means you can’t actually leave.
Open protocols. The connective tissue is where ownership gets cheapest and most durable. Two open standards now do the heavy lifting. MCP, with more than 200 server implementations, standardizes how an agent reaches tools and data. A2A, donated to the Linux Foundation, standardizes how agents built on different frameworks reach each other. Build on the open standard and your stack speaks a language no single vendor owns, the way TCP/IP made the internet indifferent to whose router carried the packet. That indifference is the asset.
The case against owning anything
Abstraction has a price, though.
The moment you route everything through a common middleware, you give up the provider-specific advantages that make the frontier worth paying for.
Prompt caching can cut costs by up to 90% on repeated context, and each lab implements it differently.
Native tool use, computer use, and extended reasoning modes are tuned per model.
Anthropic notes that the longer and more complex the task, the larger Fable 5’s lead over weaker models.
Ben Thompson has argued that the safety guardrails gating these models are the moat itself.
That argument is right, and it’s why complete model sovereignty is the wrong goal. Chasing it drags your whole stack down to whatever the most portable model can do, which is a slow way to lose to a competitor who simply uses the best tool available.
The frontier is worth renting at full price.
Also, the products these companies build around the cutting-edge models are state of the art (Codex, Claude Code, Cowork...).
You really don’t want to unplug. Own the routing and the floor, and keep the frontier on a lease. Pay for it on the work that needs it, or you can benefit greatly from caching and native features when calling a single provider.
What next
The Fable shutdown will probably be resolved; it reads like a dispute over a misunderstanding more than a permanent policy.
The lesson survives the news.
Probably the most consequential technology of our generation turned out to rest on a relationship between a company and a state that no customer was party to.
The companies that learn the lesson will treat synthetic intelligence the way they already treat power, connectivity, and capital: a critical input no single source should fully control.
If you could use some expert help in designing your AI stack for resilience and to be future-proof, get in touch with us. That is part of the work we do for our clients in the AI Labs at Visa Consulting, where we help banks, fintechs and businesses to architect and materialize their AI strategy.