Don't Marry Your AI Tool. Date a Few.

Last week, the most capable AI model in the world lasted three days before it was switched off for everyone outside the United States. No notice. No warning. Gone by Friday.

If your entire AI habit lived inside that one new tool, you'd have started Monday with nothing.

So this week, a practical question. How do you use AI in a way that survives any single tool disappearing?

It's the same instinct you already have as a QS. Don't put all your trust in one supplier with no alternative.

So don't marry your AI tool. Date a few.

What has happened

On 12 June, Anthropic confirmed that a US government directive forced it to suspend access to its two most advanced models, Fable 5 and Mythos 5. They'd been available for three days.

The order was aimed at foreign nationals, anywhere in the world. Anthropic couldn't cleanly separate its foreign users from its American ones, so to comply it switched the models off for everyone, in every country. Even its own staff who aren't US citizens were locked out.

The company says the order rests on a disputed report and that the same capability is available in rival tools. It's complying anyway, because it has no choice.

Whatever you make of the politics, the result is the same. A tool people relied on vanished overnight, on a decision none of its users had any say in.

What Fable 5 was

Worth knowing what was lost, because it wasn't a toy. Fable 5 was the most advanced model Anthropic had ever built, released alongside a sibling, Mythos 5, only days before both were pulled. For its three days in the wild, it was arguably the most capable AI on the planet.

Here's what that meant in practice.

• It went straight to first place on the independent leaderboards the industry uses to rank these tools for writing and fixing code.
• It scored close to top marks on the hardest maths benchmarks going, the sort most models fail badly.
• One well-known tracker said pulling it was the first time its best-available-AI chart had ever gone into reverse, because nothing else on the market matched it.
• One developer reported it took apart an entire 1989 computer game in a single night, mapping out work that had taken six months by hand.
• Another put together a working online multiplayer game with it in a couple of days.

Those last two are user claims, so take them with a pinch of salt. But the direction of travel was clear. This was a serious step up, and people had noticed.

Why this matters now

These tools are getting properly good, properly fast, and more of us are quietly leaning on them for real work. Drafting a compensation event narrative. Turning a month of diaries into something a Project Manager will actually read.

Picture the downside. The Project Manager wants that narrative by Friday. You've built it in your usual tool all week. Thursday morning, it loads an error page and stays there.

If that tool is the only one you know, your Friday just got a lot worse.

Leaning on AI is fine. Leaning on only one is the risk, because tools get pulled, prices change, and apps fall over on the worst possible day. The fix isn't to use less AI. It's never to let one tool become the only thing you know how to drive.

How to use AI without getting locked in

Four habits. None take long, and they make you faster anyway.

1. Keep your prompts in a document you own

Your best prompts are the real asset here, not the tool. Refine one that works, save it in a document of your own, and it'll run in anything you paste it into. Leave it buried in an app's chat history and you start from scratch the day that app is gone.

Here's one worth keeping, for turning rough notes into a compensation event narrative.

You are helping a quantity surveyor draft a factual account for an NEC4 compensation event.

Turn my rough notes into a structured narrative under these headings:

- What happened, and when (dates and sequence)
- The clause or instruction it arises from
- The effect on cost and on the programme
- The records that evidence it

Use plain, factual language. Do not invent dates, figures or facts. Where something is missing, mark it [to confirm] rather than guessing.

My notes: [paste here]

That prompt works in ChatGPT, Claude or Copilot. The tool is interchangeable. The prompt is yours.

2. Learn to drive a second model

Most QSs only really know one tool, usually ChatGPT. Spend an hour getting comfortable in a second one, like Claude or Copilot.

They're far more alike than different, so the skill transfers almost completely. The point isn't to master five tools. It's to have one you use every day and one you could switch to tomorrow without missing a beat.

3. Never paste your only copy of anything

The contract, the spec, the diaries, the photos. Keep the source in your own system. The AI is there to read and draft, not to hold your evidence.

If your only structured version of the month lives in a chat window, you've made a tool you don't control your filing cabinet. That's the trap.

4. Keep your judgement on the output

The model drafts. You sign it off. Every time.

It can draft a confident compensation event narrative in seconds and still get a date or a clause wrong. The tool can vanish overnight. Your name on the work cannot. From 9 March 2026 that oversight stops being good practice and becomes a professional requirement under the RICS standard on responsible use of AI.

Local models: the option nobody can switch off

One more option, and it's the strongest answer to everything above.

Most AI tools run on someone else's computers, which is exactly why a directive or an outage can take them away. Local models are different. They're published for anyone to download and run on their own hardware. Once a model is sitting on your firm's own kit, nobody can remotely switch it off. No directive reaches it, no price change, no outage.

The same week Fable went down, several genuinely capable models were released that anyone can download and run themselves. The term to search, if you're curious, is open weight. That's where a good chunk of the industry is heading.

The honest catch. Running one properly needs decent hardware and some setup, so it's a conversation for your IT team or supplier, not a Friday afternoon job. The best open models still trail the very top cloud ones, though the gap is closing fast.

You probably won't run one yourself tomorrow. But it's worth asking any supplier whether they can run on infrastructure you control. That's the version of AI nobody can take away.

What to watch out for

Don't spread yourself so thin you master nothing. Two tools you know well beats five you half know.

Mind what you paste in. Confidential or commercially sensitive material shouldn't go into a tool you haven't checked, because you don't always control where it lands or how long it's kept. Know your firm's policy before the contract goes in the chat box.

The same task can also give different answers in different tools, so re-check when you switch rather than trust it on sight.

Try this today

Take the prompt above, or one you already use, and run it in two different AI tools side by side. Same task, same input. See which gives you the cleaner first draft and which needs more chasing.

Ten minutes, and you've gone from one tool to two. Whichever you prefer, you've got a backup you can actually drive. If your main one disappears on a Friday, Monday already has an answer.