For three days in June, the most capable AI yet released was open to the world. Then Washington ordered it switched off for every foreign national — us included. I build with these tools for a living. This is my account of what the lockout is costing Australia, and the fix I think is sitting in plain sight.
Every minute on that clock, American firms keep building on the frontier. We don't.
By Sabour, GoldNet Engineering Group · independent · sourced throughout
Fable 5 lands at #1 on the independent Artificial Analysis Intelligence Index and is in enterprise hands within days. For three days, the new frontier is open to anyone.
For the first time ever over access to a language model, Washington orders it suspended for any foreign national, inside or outside the United States — down to Anthropic's own non-citizen staff.
It can't block foreign nationals one by one in real time, so it pulls Fable 5 and Mythos 5 for everyone on the planet. Australia loses access with the rest of the world. Other Claude models keep running.
The stated reason was a suspected "jailbreak." Anthropic called the flaws relatively simple and not universal; one report (TechCrunch) argued the ban "was never about a jailbreak" at all.1,4,9
What sets it apart is long-horizon work: it plans across stages, runs for days, keeps notes, and checks its own output before it calls a job done. Stripe used it to run a migration across a 50-million-line codebase in a day — work that would take a team months by hand. That's the difference between a tool that helps a person and one that does the project.
Multi-file migrations, deep review, security-sensitive code — run end to end, not prompt by prompt.
Stays on task across millions of tokens and improves from its own notes. The work that used to need a whole team.
#1 on the independent intelligence index at 64.9 — about 5 points clear of the field — and best-in-class on 5 of its 10 benchmarks.3
Benchmarks are the evidence, not the point. The point: this tool turns months of skilled work into days — and right now only one country can use it.3,7
The order swept up the closest US allies — the UK, Canada, Australia and New Zealand — and their citizens, with no line drawn between friend and rival. Analysts called it a blunt signal that Five Eyes partners should never have been caught by.4,5,6
A frontier lead isn't a head start you can sprint to close. Each day of exclusive use feeds four loops that make the next day's lead bigger. "We'll catch up later" assumes a gap that stands still. This one accelerates.
Work done with the best tool builds better tools, data and workflows — which do better work.
Output per worker steps up where the frontier is open, and the lead pays for the next lead.
Fluency with frontier agents is experience you can't import after the fact.
Standards, tooling and capital settle around whoever ships first. Everyone else integrates late.
The curve shows the shape of a compounding lead, not a forecast of a number. It is the mechanism, drawn — nothing measured.
This slide separates three things: the measured opportunity, an illustrative delay-exposure scenario, and the line where evidence stops and my judgement begins. I won't pretend the whole number is "lost." My point is narrower — delay has a real cost, and we almost never price it.
My assumption, not a finding: treat 10–25% of the measured opportunity as sensitive to frontier-model access and first-mover timing. It is here to make the cost of waiting visible and easy to argue with. If you have a better number, it should replace mine.
Where I stop: the economy-wide opportunity is large and measurable. The cost owed specifically to the Fable lockout is not — I can't honestly put a hard number on it. The defensible concern is delay risk, which compounds. Sectoral figures overlap the headline and are not summed.
This isn't an abstract race. We start the lockout already behind on the exact thing frontier AI moves — productivity — with a workforce still learning to use AI at all.
A country with a productivity problem just got shut out of the most powerful productivity tool ever shipped. The cost lands hardest where we're weakest — health, cyber-defence, government services, mining and energy, small-business software, research.
It was framed as one, and that deserves a straight answer. A security concern can be entirely real and the response still be the wrong tool. The honest line isn't "security was fake" — it's that a real concern is best met by controlled allied access, not a blanket door slammed on friends.
The concern was a suspected jailbreak. Anthropic called the flaws relatively simple and not universal; some reporting questioned the rationale.1,4,9
A worldwide, all-foreigners shutoff is a sledgehammer. It treats a trusted ally exactly like an adversary — the opposite of targeted.5,6
Five Eyes intelligence and AUKUS technology sharing already run on the highest bar of mutual assurance. The trust is in place.8
A bounded, auditable allied channel is more secure than all-or-nothing — because the alternative to access isn't safety. It's allies routing around the wall, and capability pooling in one country.
In US AI-chip export rules, Australia has been treated as a trusted, Tier-1-level destination alongside the other Five Eyes. That framework has since shifted — but the allied-trust principle is well established.10
AUKUS Pillar II and the 2024 ITAR reforms opened licence-free defence-technology sharing between the US, UK and Australia. Carving out trusted allies is a road already paved.8
The order simply didn't distinguish allies. That's a fixable choice — exactly the kind diplomacy resolves.5
Australia isn't asking for a favour. It's asking the US to build an AI-specific trusted-ally channel — the same logic that already underpins AUKUS defence-technology sharing, applied to model access.
A channel the US can grant with confidence because it keeps control. Three properties make it safe to say yes to:
Named models, named trusted jurisdictions (Five Eyes), defined uses. Scope is explicit, not open-ended.
Named entities, identity-verified users, access logging, incident reporting, use-case limits, no onward transfer, one Australian-government assurance point — the posture AUKUS already runs on.
The US keeps revocation rights under agreed conditions. A carve-out it can withdraw is far easier to grant than a permanent one.
Who carries it in Canberra:
On the US side it lands with the people who already run allied export carve-outs — Commerce / BIS, the State Department, the NSC, and the White House AI office.
Australia formally requests an allied access carve-out — government and security-assured industry first, then the wider builder community — through DFAT and the AUKUS channel, at ministerial level.
Don't wait for the answer. Define the bounded-verifiable-revokable model so we arrive with a ready-made way for the US to say yes.
Clear guidance so Australian firms and researchers aren't stranded while access is negotiated — and a standing position for the next model, so we're never caught flat-footed again.
Table the carve-out request. The precedent, the trust and the economic case are already here. The missing input is urgency.
Document what the lockout costs your work and add your name. A quantified, named coalition is what turns a brief into a priority.
Share it. The case gets stronger every time someone who can act sees the cost.
Australia has now been locked out for
An independent brief by Sabour, GoldNet Engineering Group · June 2026 · non-partisan. Every figure is sourced above; where a number is my own judgement (the delay-exposure band) it is labelled as such. Sector figures overlap the national headline and are not summed. Not legal or financial advice.