🔭  The moat moved. It’s not the model anymore — it’s not being trapped by one.

The fact. A single government order made the most capable model on earth vanish from millions of accounts in an afternoon. No price change, no deprecation notice — one directive, and it was gone.

The consequence. That turned an abstract worry into a line item: your access to a top model is now a dependency someone else controls. The scarce thing this week wasn’t capability — three labs ship near-frontier models. It was portability: the ability to lose your best model on a Friday and not lose your business.

The market’s answer. Sakana didn’t build a better model; it built insurance against ever depending on one. Routing isn’t a clever hack — it’s the rational response to a frontier that can be switched off by people who don’t answer to you.

The bet, with its reason. This is the part to hold me to: by the end of Q4 2026, at least one major enterprise platform ships multi-model routing as its default configuration. The forcing function isn’t fashion — it’s procurement. Once a CTO has watched a vendor go dark overnight, “single model, no fallback” becomes a continuity risk that fails a security review, the way “no backup provider” already does for payments and cloud. Lock-in stops being a default you accept and becomes a risk you have to disclose. The question shifts from “which model is best?” to “what happens to me when I lose it?”  The full case — the portability premium →

You are my AI resilience auditor. Here are my top

recurring AI-dependent tasks (start with ~5):

[For each: the model you use, what it produces, and

how often. Be specific — "I draft client proposals

in Claude, ~5/week" not "I use AI for work."]

Now stress-test it:

1. For each task, rate my lock-in risk 1–5

   (5 = only one model can do this acceptably).

2. For every task rated 4–5, name two alternative

   models that could plausibly replace it, and what

   I’d lose in the switch.

3. Flag any task where I’ve never tested an

   alternative — that’s a blind spot, not a preference.

4. Give me one migration I should test THIS WEEK

   while nothing is on fire.

End with the single task that would hurt most to

lose, and why.

🛠️  TOOL OF THE WEEK

OpenRouter

One API, every major model — switch between Claude, GPT, Gemini and a hundred others without rewriting a line of code.

★★★★½ / 5

The cheapest insurance against this week happening to you. OpenRouter sits between your app and the model providers: you write to one endpoint, and route requests to whichever model you like — or automatically fail over when one goes dark. The Fable blackout is exactly the scenario it’s built for. If your workflow had been pointed at OpenRouter on 12 June, swapping to Opus 4.8 or Gemini would have been a one-line config change, not a fire drill.

Not a builder? The principle still applies: keep a second tool you know how to use, so a blackout costs you minutes, not a deadline. The version below is for the API crowd.

Use if: you build anything on AI APIs and can’t afford a single vendor deciding your uptime. Skip if: you only use one chat app and never touch an API. Describe it to a colleague: “It’s the universal adapter for AI models — one plug, every socket.” Best use case: wire your most critical workflow through it now, with a named fallback model, before you need one.  OpenRouter →

💡  TIP OF THE WEEK

Never name “latest” in anything automated

Here’s where most people get burned silently: in any script, API call or automation, pin the exact model version — never “latest,” never the default. The reason landed hard this week. Workflows pointed at Fable 5 didn’t get a warning; they got an error, mid-run, when the model vanished. But the quieter danger is the opposite — a model silently upgrading under you and changing its behaviour overnight.

Pin exact version strings. claude-opus-4-8, not “the newest Claude.” Silent upgrades break things that worked yesterday.

Name a fallback model in code. When your primary errors out, route to a named second choice automatically — don’t let a 5:21pm directive take you offline.

Keep one alternative warm. Run a real task through a backup model monthly. An untested fallback isn’t a fallback; it’s a hope.

Why it works: pinning turns “the AI broke and I don’t know why” into a single line you can read and change. Where it doesn’t apply: in casual chat use, “latest” is fine — let it upgrade. The discipline is only for anything that runs without you watching.

For the colleague whose entire workflow runs on one model — and who’s never thought about what happens if it goes dark.