Claude is back for Australian businesses. Don't just plug it back in

Three weeks after switching Australian businesses off, the US has switched them back on. As Startup Daily reports, the US Commerce Department lifted its export controls on Anthropic’s frontier models on 30 June, and access to Claude Fable 5 and Mythos 5 started returning to international customers, Australians included, from 1 July.

If your business felt the June outage, this is the news you were waiting for. It is also worth reading the fine print before you celebrate.

Back, but with strings attached

The controls were not lifted because Washington changed its mind about the risk. They were lifted because Anthropic agreed to conditions: proactively detecting and blocking jailbreak techniques, and reporting malicious activity to the US government. The government has explicitly reserved the right to reimpose restrictions if those commitments slip.

In other words, access did not go back to normal. It went from “off” to “on until further notice”. The switch still exists, it still sits in another country, and it has now been flicked twice in a month. That is not a reason to avoid these models. It is a reason to treat them as what they are: a supplier that can be interrupted, not a utility that can’t.

The outage was a free audit. Don’t waste it

When the ban landed in June, we wrote about what it exposed: businesses that had wired a single US-hosted model into the middle of their operations found out exactly how load-bearing it was, with no notice and no recourse. That advice has not changed now that access is back. What has changed is that you have a rare, concrete data point about your own exposure.

For three weeks, you got to see precisely which parts of your operation degrade when one provider goes dark. Before the memory fades, write it down:

1. What actually broke? List the workflows that stopped or got worse during the outage. That list, not a theoretical risk register, is your real dependency map.
2. What carried on? Anything that kept working either didn’t depend on the model or had a fallback. Notice what made those workflows resilient.
3. What did the workaround cost? If staff picked up the slack by hand, roughly cost out those hours. That number is your budget case for fixing the single point of failure properly.
4. Switch back deliberately. As access returns, restore your highest-value workflows first and confirm each one works, rather than flicking everything on at once and assuming.

You probably don’t need sovereign AI. You need portable AI

The episode has reignited the sovereignty debate. In SmartCompany’s Neural Notes, local AI figures argue Australia needs models and infrastructure under domestic control, because you can’t have agency while someone in another country can turn off the tap.

That debate matters at a national level. For a Melbourne service business, the practical version is smaller and cheaper. You do not need a sovereign model to run your quoting, enquiry triage or follow-up automation. You need those workflows built so the model behind them is a component you can swap, not a foundation you are welded to. The workflow logic (“classify this enquiry, draft a reply, update the CRM”) does not care which model does the work. Build it that way and a future ban, price hike or outage becomes a configuration change instead of a crisis. We covered the same principle from another angle in putting AI in your CRM, not just ChatGPT: the value lives in the workflow, not the model.

The func.digital take

Use the best model available to you, and right now that includes the frontier Claude models again. Just carry the June lesson forward: the outage told you exactly where your operation is fragile, and access came back conditional, not guaranteed.

If you are not sure how exposed you are, or the honest answer to “could we swap models in a day?” is no, that is exactly what a free digital systems audit is for. We map where a single provider sits on your critical path and what it takes to make that swappable, before the switch gets flicked a third time.