America Says Don’t Bother. An Australian AI Model Already Beat OpenAI.

I wrote a three-part series about whether we should trust Chinese AI. Then America blocked everyone else from using its best AI. Now I’m wondering — what if Australia just built its own? Not as an expert. As a small business owner who uses AI every day and wants to know if the Aussie stuff is actually any good. Spoiler. One of them already beat OpenAI.

This is Part 4 of The China Question series. Parts 1 to 3 asked whether we should trust Chinese AI models like Kimi and DeepSeek. The answer was complicated. They’re cheap, they’re good, and they’re controlled by a government that doesn’t share our values.

But here’s the thing. In June 2026, America did something that made the Chinese question look almost academic. The US government forced Anthropic to cut off access to its most powerful AI models — Mythos 5 and Fable 5 — for everyone outside the United States. No notice. No consultation. Just a switch flipped overnight.

Two weeks earlier, Australia’s own national security agencies had been granted access to Mythos. Two weeks later, it was gone. The ban was eventually lifted in early July. But the message was clear. A foreign government can cut off your AI supply overnight. And there’s nothing you can do about it.

Then OpenAI delayed the global release of GPT-5.6 at the US government’s request. Only about 20 government-vetted US organisations got access initially. The rest of the world waited.

So if China is one risk and America is another risk, what’s Australia doing about it? I’ve been digging into this for weeks. Here it is, in plain English, for people who run businesses and use AI tools but don’t have a computer science degree.

Think of it like a car

Before we get into the Australian AI stuff, let me explain how AI actually works — in language a small business owner can understand. There are three parts to any AI tool you use.

The Brain

The brain is the AI model itself. Claude. ChatGPT. Gemini. DeepSeek. These are the brains. They’ve been trained on massive amounts of text and data so they can think, write, code, and reason. When you type a question into ChatGPT, the brain is what figures out the answer.

Right now, almost every brain Australian businesses use comes from America or China. OpenAI (American). Anthropic (American). Google (American). DeepSeek (Chinese). Moonshot/Kimi (Chinese). We rent these brains by the word — technically, by the “token,” which is roughly three-quarters of a word.

The Engine

The engine is the hardware that runs the brain. Specialised computer chips called GPUs. These chips are incredibly expensive and incredibly power-hungry. You can’t run a serious AI brain on a laptop. You need a data centre full of GPUs.

Right now, almost all the engines are overseas too. American data centres for American brains. Chinese data centres for Chinese brains. When you use Claude, your data travels to America, the brain thinks, and the answer travels back. Every single time.

The Agent

The agent is what you actually interact with. The chatbot on a website. The email writer in your CRM. The tool that answers customer questions at 2am. The agent uses a brain to do work for you. You don’t see the brain or the engine — you just see the agent doing its job.

I build agents for Australian businesses. Kween Chat is an agent. It uses a brain (currently Claude or ChatGPT) running on an engine (American servers) to help my clients sell things 24/7.

The question is. Can we build Australian brains? Can we run them on Australian engines? And are they any good?

America says “don’t bother, Aussies”

In July 2026, the United States Studies Centre at the University of Sydney published an article with a blunt message. Australia shouldn’t try to build its own frontier AI.

Their argument, written by Olivia Shen, Director of Strategic Technologies at the USSC, is simple. Training a world-class AI model will cost over US$1 billion by 2027. No country outside the US and China has managed it, despite billions spent by France, the UAE, Japan, South Korea, India and others. Australia doesn’t have the capital, the talent pool, or the computing power to compete at that level.

They say we should focus on what we’re good at. Critical minerals (which go into chips). Data centre infrastructure (which houses the engines). Specialised datasets (which train the brains). Be the supplier, not the competitor.

It’s a fair argument. It’s also the argument that keeps you permanently dependent on someone else’s brain and someone else’s engine. As we saw in June, that someone else can cut you off without warning.

What Australia is actually doing (three things)

1. The government is building the rules

On 15 July 2026, Prime Minister Albanese announced a world-first national AI framework. A new Office of AI inside the Department of Prime Minister and Cabinet. Australian Standards for AI — the first to be legislated by any government worldwide. Copyright protections for Australian artists and writers. Rules for data centres to underwrite their own power supply and be water-efficient.

Behind that, the National AI Plan released in December 2025 commits over AU$362 million in targeted grants. The APS AI Plan is building the GovAI platform — secure AI tools for government, running entirely inside Australian Government infrastructure.

The government is also funding the AI Safety Institute. They’re not just talking. They’re spending real money and writing real rules.

2. An Australian brain already beat the world

This is the part that surprised me most. In October 2025, an Australian startup called Isaacus released a model called Kanon 2 Embedder. It ranked first globally on the Massive Legal Embedding Benchmark — the world’s standard test for legal AI.

Not second. Not tenth. First. Above OpenAI. Above Google. The official benchmark results show it was 9 percent more accurate than OpenAI’s Text Embedding 3 Large and 6 percent more accurate than Google Gemini — while running 340 percent faster.

It’s a specialist brain, not a general one. It does legal document search and classification, not creative writing or chat. It’s available now on AWS Marketplace at about US$1 per million tokens. But it proves something critical. Australian AI can beat the world’s best when it’s built for a specific job.

3. Australia is building its own general-purpose brain

A Sydney company called Sovereign Australia AI launched in September 2025 with 256 Nvidia Blackwell B200 GPUs — the largest GPU deployment in Australian history. AU$100 million in funding. Hosted at NEXTDC data centres. Fully onshore. The Australian Financial Review covered the launch — the founders say they can build Australia’s ChatGPT for under $100 million.

They’re building two brains. Ginan — a free open-source model for researchers. Australis — a commercial model for businesses and government. Both designed to understand Australia natively. Our laws (ASIC, APRA). Our language. Our regulatory frameworks. Not an American brain pretending to understand Australia. An Australian brain that actually does.

The founders have serious credentials. CEO Simon Kriss has advised the United Nations on AI strategy. Co-founder Troy Neilson was head of data and AI at Telstra Purple. These aren’t amateurs.

Although honestly, I wish they’d change the name. “Sovereign Australia AI” sounds like something a far-right politician would launch at a press conference surrounded by flags. I’m not saying the founders are like that — they’re clearly serious tech people doing serious work. But every time I type the name, I feel like I’m about to start ranting about sovereignty and borders and taking back control. It’s giving Brexit energy. Maybe just call it “Aussie Brain” or “G’day GPT” or literally anything that doesn’t sound like a political movement. Please. I’m begging.

The models aren’t commercially available yet. They’re still being trained. But the infrastructure is real, the funding is committed, and the GPUs are in the country.

The engines you can use right now

While Australis is still cooking, there are two Australian platforms running AI brains on Australian engines today. For small business owners, this is where it gets practical.

SCX.ai (Southern Cross AI) — the most mature option

SCX.ai launched Australia’s first sovereign AI infrastructure node in Sydney in January 2026, in partnership with SambaNova Systems. They’re now raising AU$40 million through an IPO on the ASX — so they’re a serious, publicly-listed Australian company.

They have their own Australian brain called MAGPiE — a model built specifically for Australian context. It understands Australian law, Australian regulations, and Australian cultural patterns natively. Not through prompting. Through training. They also host other models (DeepSeek, Llama, Qwen) on Australian infrastructure.

In car terms. SCX.ai builds the engine, houses the brain, and lets you drive. The engine is in Australia. The brain never leaves the country. Your data stays here.

Pricing. AUD $3 to $4 per million tokens. For comparison, Anthropic charges $15 to $25. OpenAI charges $10 to $30. Same capability, Australian-hosted, roughly 80 percent cheaper.

The API works with existing code. If you’re using the OpenAI or Anthropic API, you change one line — your base URL — and everything else works. No rewrites.

Continuum Labs — the budget option

Continuum Labs is a smaller Australian AI research company. They host open-weight models (mainly DeepSeek V4-Flash) on Australian infrastructure. Same price — AUD $3 to $4 per million tokens. OpenAI-compatible API. One line to switch.

They’re not building their own Australian brain. They’re running existing open-weight brains on Australian engines. If you just want good AI at a fraction of the cost with data that stays in Australia, this is the simplest option.

Their hosting meets ISO 27001 and SOC 2 compliance standards. For small businesses that care about data security but don’t need a custom-built Australian model, Continuum is the straightforward choice.

How this connects to The China Question

In Part 1 of this series, I asked whether we should trust Chinese AI. The answer was — it’s cheap, it’s good, but you’re sending your data to a country with different values and different laws.

In Part 2, I looked at the economics. Chinese models like DeepSeek and Kimi are dramatically cheaper than American models. But “free” from a Chinese company comes with strings you can’t see.

In Part 3, I explored the security angle. When your AI brain is hosted in another country, your data leaves your control. Full stop.

Now in Part 4, the question flips. If Chinese brains are risky and American brains can be switched off — what about Australian brains on Australian engines?

The answer is. We’re not there yet for general-purpose AI. But we’re closer than the “don’t bother” crowd wants to admit. And for some jobs, we’re already winning.

But is it actually good? (The honest bit)

Here’s where I have to be straight with you. For general-purpose chat — writing emails, answering customer questions, generating marketing copy — the open-weight models on SCX and Continuum are good but not quite as good as Claude or ChatGPT at their best.

According to Continuum Labs’ own benchmarks, DeepSeek V4-Flash delivers 85 to 95 percent of frontier closed-model capability. That means for most everyday tasks, you won’t notice the difference. For highly complex reasoning or creative writing, you might.

For specialised Australian tasks — legal document processing, regulatory compliance, understanding Australian business context — the Australian models are already competitive. Kanon 2 proved that. MAGPiE is built for exactly this.

Think of it this way. If you need a brain surgeon, you want the best brain surgeon in the world. If you need a GP who knows your medical history and understands the local healthcare system, the best GP in your suburb might actually be better than a brain surgeon in another country.

Same with AI. For general tasks, the American brains still lead. For Australian-specific tasks, the Australian brains are already matching or beating them. And the gap is closing fast.

What this means for your small business

You’re probably not going to build your own AI model. But you can choose where your AI runs and whose brain it uses. Here’s the practical takeaway.

• If cost matters — Australian platforms are 80 percent cheaper than American ones. Same jobs, same quality for everyday tasks, fraction of the price.
• If data sovereignty matters — Australian platforms keep your data in Australia. It never crosses a border. No foreign government can cut it off.
• If you’re in a regulated industry — legal, financial, government, healthcare — Australian-hosted AI meets APRA CPS 234 and Privacy Act requirements out of the box. No cross-border data flow assessments needed.
• If you want to support Australian innovation — SCX.ai is about to list on the ASX. Sovereign Australia AI is building frontier models. Isaacus already beat the world. These are Australian companies doing genuinely world-class work.

Why I’m keeping this on my radar

I build AI agents for Australian businesses. If I can offer a client an AI sales agent that runs on Australian infrastructure, at 80 percent less cost, with data that never leaves the country — that’s not patriotism. That’s a competitive advantage.

The US Studies Centre says don’t bother building frontier AI. Fair enough. But SCX.ai isn’t trying to build Claude. They’re running good models on Australian engines at a fraction of the price. Sovereign Australia AI isn’t trying to beat GPT-5. They’re building a brain that understands Australia natively. Isaacus already beat OpenAI in their lane.

Maybe the answer isn’t “build a better ChatGPT.” Maybe it’s “build the right brain for the right market and run it where your customers actually live.”

I’ll be testing SCX.ai’s API against Claude for my own use cases over the coming weeks. Same prompts. Same tasks. Same evaluation criteria. If it holds up, I’ll switch. Not because I’m sentimental about Australian AI. Because it’ll be cheaper, sovereign, and good enough.

Watch this space.

Want an AI agent built for your business? Let’s talk →

Frequently asked questions

This post contains no affiliate links. It is an independent research piece based on publicly available information and official government sources. All claims are linked to their original sources. Pricing and availability are current as of July 2026 and may change. Always evaluate AI models against your specific use case before switching providers.