Don't build your practice on one AI model

Update — October 10, 2026. The example this article opened with has now played out in full. After the June export-control restriction came a sunset date, two extensions, a permanent split by plan in July, and then a successor: Fable 5.1 in September, with Opus 5.5 close behind. A setup that lived in files would have needed nothing rebuilt at any of those steps, which is the argument below. The current lineup is on which model you get on which plan.

In June, the advice everywhere — ours included — was that Claude Fable 5 moved the line on what you could hand an AI in one shot. Days later the line moved again: the US government placed export controls on the model, Anthropic restricted access to US nationals, and a lot of people outside the US opened their model picker to find the tool they'd just reorganized their week around was gone. That restriction was later lifted, and the access story was rewritten several more times (see the note above). It was a clean, expensive lesson, and worth learning whether or not any one of those changes touched you.

The lesson goes beyond Fable 5: you don't own the model. You rent access to it, on terms a company and now a government can change without asking you — and a practice wired tightly to one specific model is one announcement away from a bad Monday.

You've always been renting

This is easy to forget when the tool is reliable, so say it plainly: every AI model you use is a service someone else operates, and the terms are theirs. They can deprecate it — models get retired, and the replacement behaves differently. They can reprice it. They can throttle it under load, or change the usage limits on your plan. It can go down for an afternoon during the one week you're slammed. And — as Fable 5 has now demonstrated repeatedly in a single summer — access can be cut off by geography or nationality with little notice, or a model can be pulled from subscription plans onto separate usage-based billing on a few weeks' warning, with the deadline itself moving twice along the way.

None of that is a reason to use AI less. It's a reason to be deliberate about what you let depend on it. There's a difference between "this model makes my Tuesday faster" and "I cannot produce a proposal without this specific model in my picker." The first is leverage. The second is a single point of failure you built yourself, and the Fable 5 restriction showed how fast that bill comes due.

What you actually own

Here's the reassuring half, and it's the more important one. The model is the most replaceable part of your setup. Everything that makes AI useful for your practice specifically is the part you own, and none of it cares which model is behind the picker.

You own the judgment about what to hand over and what to keep — the call that some work is delegable, some is augmentable, and some never leaves your hands. That decision is model-independent by definition; it's about the nature of the work, not the engine.

You own the context: who you serve and who you don't, your offers and your rate floor, how you sound, the state of each live engagement. When those live in files you control, any model can read them. A new model on Monday inherits your whole setup by lunch.

You own the briefs and the standards: the objective-inputs-constraints-done-why structure that equips a model to produce a real deliverable, and the SOPs that define what good output looks like. A brief written to the top capability tier runs on whatever model currently occupies that tier. It does not have a model's name baked into it.

The model is the engine. Your operating model and your setup are the car. Engines get recalled; you keep the car.

The over-fitting trap

The way practices get fragile is subtle, and it usually looks like diligence. You find a model you like. You tune your prompts to its quirks — this phrasing gets better output, that one trips it up. You lean on its specific memory feature, its specific file handling, its specific voice. Six months in, your workflow is a cast of one particular model's behavior, and you've mistaken "good at using this model" for "good at using AI."

We made this exact argument about prompts when Fable 5 shipped: the elaborate, over-specified prompt libraries people built were scaffolding for weaker models, and a more capable model made them worse, not better. Over-fitting to a model is the same mistake one level up. The fix is the same too: invest in the parts that transfer, keep the model-specific tuning thin and disposable.

The portability test

One question tells you how exposed you are. If your top model vanished from your picker tomorrow, what would actually break?

If the honest answer is "my prompts stop working" or "I'd have to rebuild my whole workflow," you've over-fit, and you should move weight onto the parts that travel. If the answer is "the same work would be a little slower or need an extra pass on a lesser model" — that's resilience. That's the state to build toward. Not independence from AI; graceful degradation when any single model goes away.

Run the test against the Fable 5 upheaval specifically. If you'd built your delegation around it, the recovery isn't a rebuild — it's routing the long-leash, whole-deliverable work to the next tier down and reading hard on the output. We wrote the model-by-task version of this before the restriction landed, and it holds up unchanged: the rule was always to match the model to the task, with named tiers that swap as the lineup changes. If you picked by tier, the floor barely moved. If you picked by name, you had some rebuilding to do.

What to actually do

Four moves, none of them urgent, all of them the kind of thing you do once and benefit from on every future model release and every future restriction.

Keep your context in files you control. Your ICP, your offers and rate floor, your voice, one file per active client. Not locked inside one tool's proprietary memory — plain documents you can paste into anything. This is the most useful single step, because it makes switching models a copy-paste instead of a migration. (If you'd rather start from finished templates than a blank page, the Claude Project Kit ships exactly these files, and they load into any model that reads documents.)

Write briefs to the tier, not the model. "Hand this to the most capable model I can reach" ages well. "Hand this to Fable 5" was out of date within weeks. The five-part brief — objective, inputs, constraints, what done looks like, why — is model-agnostic on purpose. Keep it that way.

Know your fallback and keep defaults that degrade gracefully. You should always be able to name your second choice for each kind of work. If your top tier disappears — or moves to a plan you don't pay for — the routing shifts down a notch and you read the output harder. Not a crisis, a known procedure. The fallback tier has been excellent for a while now, and each new release makes it more so, which is a big part of why this approach keeps working.

Don't reorganize your business around a model in its first month. Try new models eagerly; rewire slowly. A capability is worth building a workflow around once it's proven stable and available to you — not in the news-cycle window when its access terms are still settling. Committing hardest and fastest to a three-week-old model is what made the Fable 5 restriction so expensive.

The frame underneath all of this

All of this comes back to one idea: the asset that lasts in an AI-assisted practice is your operating model — what you delegate, what you keep, how you brief the difference — not any tool that happens to execute it this quarter. That's the whole argument of The Solo Consultant's AI Playbook, and it's deliberately model-agnostic for the reason this summer demonstrated: the tools change under you, sometimes overnight, sometimes by government order. The practice that's built on what you own survives all of it.

Fable 5 did come back, on new terms, and within weeks a newer model replaced it at the top. The next change will arrive the same way, on someone else's schedule. Build the setup that makes that matter less, and the next time a model arrives or leaves, you'll barely have to look up.