Battle of the Claws: the winning move in AI agent war is the one no one is selling

The week after that, both Meta and Google were caught racing to ship consumer clones of the very tool that did the deleting.

The week after that, Apple's chief executive Tim Cook went on the company's Q2 2026 earnings call and confirmed that the Mac Mini and Mac Studio are sold out worldwide because of demand from people running AI on their own hardware, and may remain so for months.

And in this month's Federal Budget, the Treasurer made the $20,000 instant asset write-off a permanent feature of the tax system, removing the deadline pressure that small businesses have planned around for a decade.

That is roughly the last 60 days in agentic AI. If you run an Australian business, the side of this debate the press is selling you is the wrong one.

The operators who already know that are out there trying to buy the hardware while it is still findable. This article was scheduled for publication in late April but it has been rewritten three times since.

Each rewrite was triggered by a public move from one of the companies involved: a restriction announced, a restriction walked back, a new product launched at the exact buyer this article was addressed to.

The Battle of the Claws turned out to be a story I could not finish writing fast enough. The pace at which the order of the race is changing, and who is leading it, has itself become part of the argument.

That is also part of why I think most Australian SMEs need to make their move sooner rather than later. Over the past fortnight in this masthead, we have walked through what AI agents are and how they differ from chatbots, and compared the six commercial and open-source agents now competing for Australian SME budgets.

This piece is the opinion follow-on to those two: what I think most Australian SMEs should actually do, and why the next four weeks may matter more than the next four years.

For readers joining this thread for the first time, this is the short version of what has happened: the dominant open-source agent, OpenClaw, became the most-starred software project in GitHub history in roughly 60 days.

Anthropic initially responded by restricting its tools from working with OpenClaw on standard subscriptions, then four months later walked the restriction back and launched a small business product targeting the same buyers.

OpenAI responded by hiring OpenClaw's founder, redirecting resources into its Codex platform under competitive pressure from Anthropic, and explicitly endorsing the opposite posture on third-party tool support. Meta and Google were separately reported to be building consumer clones, called Hatch and Remy respectively.

Apple's CEO has confirmed on a publicly-reported earnings call that the Mac Minis and Mac Studios people are using to run all of this are sold out worldwide and likely to remain so for months.

The Australian dimension of that supply story is real; I have spent a fortnight visiting Apple flagship retail, three JB Hi-Fi stores, and Harvey Norman trying to buy one. None had stock.

That is the news. Now the argument.

Five of the six are American - the rules they play by are being rewritten quarterly

The six players competing to run agentic work inside Australian businesses break down on a single axis that almost no commercial AI coverage discusses honestly. 

Five of the six are American technology companies. Four of those five (Anthropic, OpenAI, Meta, Google) are now competing with each other on the terms of access, the cost of usage, and the openness of their platforms, in real time, in response to competitive pressure that small business owners cannot see and cannot influence.

The rules are being rewritten quarterly. Sometimes monthly. Consider what has happened in this single calendar year so far.

In early 2026, Anthropic restricted standard subscription access to third-party agents and dropped one of its team-tier products by a third with no negotiation.

We covered both moves in the earlier pieces in this series, framed at the time as evidence of a closed-garden commercial posture.

Then in April, Reuters reported that pressure from Anthropic's Claude Code product had led OpenAI to redirect resources toward Codex, OpenAI's developer-first agent platform, explicitly competing on openness, third-party integration, and a Code-friendly Agent SDK.

Within weeks, OpenAI had Codex integrated with GitHub, Figma, JetBrains, Xcode, a Chrome extension, a mobile control surface, and an enterprise plugin marketplace.

Codex reportedly passed two million weekly active users. Then Anthropic walked back the restriction. From 15 June 2026, Max subscribers receive a $100 monthly credit specifically for Agent SDK and third-party tool usage, exactly the workflows the earlier restriction had constrained.

Anthropic's framing is that the previous rules had been "unclear" and the change is a "clarification." That is the polite version.

The substantive version is that the most principled of the commercial group restricted, faced competitive pressure from a more open competitor, and reversed itself inside four months.

Then Anthropic launched Claude for Small Business, a packaged product placing Claude inside Intuit QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365, with 15 pre-built workflows specifically targeting the small business segment.

The launch quote from Anthropic's president names "small businesses that have never had the resources of bigger companies" as the target market.

Take a moment with that sequence. A vendor restricted developer-tool access, competitive pressure from a rival pushed them in the opposite direction, they reversed the restriction, and within 24 hours of the reversal taking effect they had launched a packaged product targeting the exact buyer this series of articles is written for.

Four moves, four months. This is not a stable commercial landscape. This is six well-capitalised companies fighting a high-speed strategic war for distribution into Australian small businesses, and Australian small businesses are the territory being fought over.

Anthropic is not the villain in this story. Anthropic is, on the available evidence, the most responsive to feedback of the commercial group.

The point is not that any one of the four is acting badly. The point is that every one of them is doing what well-capitalised commercial competitors do under pressure: changing the rules, changing the prices, changing the products, changing the access terms, in response to forces that are entirely invisible to the Australian SME owner who just wanted a stable platform to build a few workflows on.

This is the cloud-versus-on-premise decision of 2012, replayed. Most Australian SMEs went cloud, regretted parts of it, and ended up paying considerably more than they had budgeted.

The lesson, that convenience is a real cost rather than a free benefit, is being offered again. This time the stakes are higher, because the agent layer touches everything: your calendar, your inbox, your accounting, your client files, your contracts. And the agent layer is being rewritten in public, in front of you, every few weeks.

My position

For most Australian SMEs, the right answer over the next 12 months is some version of the following: own your hardware, own your orchestration layer, run a competent free AI model locally for the 95 per cent of work that doesn't need frontier intelligence, and pay-as-you-go for the 5 per cent that does.

The case is not ideological. It is practical, and it is now. Local AI models have caught up with their commercial cousins for most ordinary office work.

The free models you can run on hardware you own (Qwen, Llama, Gemma, the names are unfamiliar but the capability is real) have closed the meaningful gap with the paid commercial alternatives during 2025 and into 2026.

They are not better than the very best frontier models, and for hard problems you should still use the best frontier models. But for drafting, summarising, reconciling, sorting, classifying, and the long tail of office tasks that fill a working day, they are good enough.

The economics are not subtle. A five-person team running on commercial agent subscriptions over three years costs in the order of $50,000. The same team running on a Mac Mini and capped frontier-API calls costs roughly a tenth of that, with hardware that retains residual value at the end of the period.

We laid out the maths in detail in the previous piece in this thread; the headline is that the difference is not a rounding error, it is a small business car.

A locally-hosted model also offers something no cloud provider can match: it can hold the entire history of your business in context simultaneously - every contract, every email thread, every invoice.

A merely competent AI that has read everything your business has ever produced will outperform a frontier AI that has read three carefully-selected paragraphs.

This is the part the cloud vendors do not want SME owners to understand. Their entire business model depends on you accepting that AI is something you rent in small slices, not something you own that knows your business completely.

Sovereignty is no longer abstract, either. Privacy Act reforms, the OAIC's notifiable data breach scheme, and the looming 2026 amendments are all moving in one direction: the legal cost of having your client data sit on offshore servers is going up, not down.

A model running on a machine in your own office is the cleanest answer to that question.

The remaining 5 per cent of work, the genuinely hard reasoning where frontier intelligence is worth paying for, is best handled by a single capped API account with whichever commercial provider currently offers the best mix of capability and openness.

On current evidence I would lean OpenAI, given their public posture of supporting third-party agents. Pay per token, not per seat.

Treat the frontier labs the way a manufacturer treats a specialist sub-contractor: useful for the hard parts, not the operating system.

'But Anthropic just launched a small business product'

On May 13, Anthropic launched Claude for Small Business: a packaged bundle of Claude inside Intuit QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365, with 15 pre-built workflows covering payroll planning, monthly close, campaign drafting, invoice chasing, and contract review.

The product is good. The integrations are real. The Anthropic engineering and partnerships work behind it is genuinely impressive.

It will be right for some Australian small businesses. Specifically, it will be right for businesses that have no internal IT capacity, are already deeply embedded in the QuickBooks-PayPal-HubSpot stack, have no specific data sovereignty requirements, have no client data sitting under professional confidentiality obligations, and are happy to operate inside whatever commercial terms Anthropic offers in 12 months, in 24 months, and in 36 months.

For that profile of business, Claude for Small Business is probably the easiest answer, and the easiest answer has real value.

I would not talk that business owner out of it. For everyone else, the local-first answer remains the right one, and the launch of Claude for Small Business is itself an argument for it rather than against it.

There are three reasons for this. The first is that a packaged commercial product targeting your buyer is, in this market, the moment at which platform lock-in tightens fastest.

The 15 workflows ship with Anthropic-defined assumptions, Anthropic-managed connectors, and Anthropic-controlled commercial terms. They are sold to businesses on the basis that they save time today.

They will, in 18 or 24 months, become the thing those businesses are reluctant to migrate away from when the pricing or the terms change.

Anthropic has done nothing wrong here. This is the playbook every successful cloud SaaS company has run for 15 years.

What has changed is only that the layer being captured is no longer your CRM or your accounting. It is the layer that talks to your CRM and your accounting and your inbox and your contracts.

The second is that the same workflows are buildable on a local stack. Payroll reconciliation between QuickBooks and PayPal, monthly close packets, campaign drafting from HubSpot data, invoice chasing, contract review: these are not exotic capabilities.

They are agentic workflows that a competent local model running through an open orchestration layer can be configured to handle, with data that never leaves a machine in your office, against a hardware cost that gets written off in the year of purchase and runs for three to five years.

The capability gap between "Anthropic-packaged on QuickBooks" and "locally orchestrated against the same QuickBooks API" is real but small, and closing.  The sovereignty and cost gap is large, and widening.

The third is that the launch of Claude for Small Business, viewed alongside the restriction in early 2026 and the walk-back this week, completes a sequence that should give any operator pause.

Four moves in four months, all in response to competitive pressures the buyer cannot see. Whatever the right product looks like for an Australian SME a year from now, it will not be exactly the product Anthropic shipped yesterday.

The version available in 12 months will have different pricing, different integrations, different commercial terms, and almost certainly different competitive positioning relative to whatever OpenAI, Meta, and Google have shipped in the same window.

The most expensive mistake at this moment is to lock workflows into a packaged product whose terms are guaranteed to move.

The simpler way to put this is that Anthropic has built a good product for the small business that wants to be told what to do.

The argument of this article is for the small business that prefers to know what is going on, and to keep the keys to its own operation while the giants fight it out.

The deadline is gone. The hardware is going.

For most of this calendar year, the urgency case for acting on local AI hardware rested on three independent timing pressures: a hardware shortage that was getting worse, a wave of consumer-AI launches expected to make it worse still, and a $20,000 instant asset write-off legislated to expire on 30 June.

As of this month's Federal Budget, one of those three has changed. The Treasurer has confirmed that the instant asset write-off is being made a permanent feature of the tax system, ending a decade of year by-year extensions that have created planning uncertainty for small businesses since 2015.

The $20,000 instant asset write-off will become a permanent feature of the tax system following this year's Budget.

For small businesses with aggregated turnover under $10 million, that removes the year-by-year uncertainty that has dogged this measure since 2015.

Genuine business hardware purchases under $20,000 - and that includes most AI workstation configurations - can now be deducted in full in the year they're installed and ready to use, with no end-of-financial-year cliff to plan around.

For operators who have been waiting to see whether the threshold would survive, the answer is now yes. The only remaining consideration is whether the hardware itself is available.

That last sentence is the whole argument. The tax window is no longer the binding constraint. The supply window is.

Apple's CEO has confirmed on the company's own earnings call that the Mac Mini and Mac Studio "may take several months to reach supply demand balance".

Higher-memory configurations of both have been pulled from the Apple Store entirely.

Lower-RAM units that remain listed are quoting four-month shipping delays. EBay scalpers are listing units atA$1,100 to $1,500 against a $999 RRP.

The cause is twofold and structural: a global DRAM shortage driven by AI infrastructure demand, and unexpectedly strong consumer-and-SME demand for unified-memory desktops to run local AI.

Neither pressure is going to ease in the next quarter. Cook said as much.

The maths is no longer about beating a deadline. It is about whether the hardware you need is on a shelf at the moment you decide to buy.

Bought now, a Mac Mini comes off your taxable income in full and starts compounding returns inside your operation immediately.

Waited on, it may not be possible to buy at any price until late 2026.

The deadline is gone. The hardware is going.

What I expect to happen

The consumer-AI launches from Google and Meta are now closer than they were when this series began. Within weeks, Remy or Hatch in some form is likely to be available to consumers, free, integrated with the platforms most Australians already use, and difficult to ignore.

By the second half of this year, every business publication in this country will likely be running explainers about how to "supercharge your team" with one or the other. I will not be one of them.

The polished, free, deeply-integrated consumer agent is the product whose price comes due 18 months later, when the workflows are too embedded to leave.

We've seen this film before. The protagonist is always the small business that locked its data into someone else's platform because it was easier than the alternative.

What could prove this argument wrong? If frontier AI capability continues pulling away from open-source faster than open-source can close the gap (which is plausible, and which Anthropic, OpenAI, and Google are all spending tens of billions of dollars to make true), then the 95 per cent local, 5 per cent frontier ratio drifts towards 60/40 within 18 months and the economics shift.

I may be calling a market direction that does not arrive in the form I expect. The hedge against being wrong is the same as the case for being right.

Don't lock yourself in. Pilot the open path, run it parallel to a small commercial subscription, keep both options live until one obviously wins.

What I'd actually do

For a typical Australian SME of around 20 staff, the practical playbook for this quarter is straightforward.

Buy a Mac Mini while you still can, even an entry-level configuration; it is enough to run the orchestration layer itself, with compute added later by networking additional machines or paying per call to a frontier API.

Pilot OpenClaw or NemoClaw on a single workflow such as invoice triage or inbox sorting before committing to anything bigger.

Run a free local model for everything that doesn't need frontier reasoning.

Hold a single, capped API budget with one commercial provider for the small number of tasks that justify the spend. And treat any vendor's terms as subject to change quarterly, because Anthropic just demonstrated, in four moves over four months, that they are.

The alternative this time is, for now, genuinely accessible. The hardware is on the shelf, just barely. The software is open source, free, and the most-starred project on the internet.

The local AI is good enough for the work most Australian SMEs actually need to do today. The economics are obvious once the numbers are on a page. And the tax treatment is now permanent.

The decision to act is no longer constrained by an end-of-financial-year clock. It is constrained only by whether what you need to buy is still available the next time you check.

The Battle of the Claws is being fought between four offshore giants and an open-source weekend project.

The Australian SME's job is not to pick a side in that war. It is to stay out of it long enough that, whoever wins, you still own the keys to your own business.

Luke Vaughan is the publisher of Business News Australia and an Australian entrepreneur recognised by this masthead in its Top 100 Young Entrepreneurs before acquiring it. Through Vaun Group, he has founded and oversees the operation of several Australian businesses, and is currently rolling out the architecture described above across the group rather than recommending it from the sidelines.

A footnote, in the spirit of full disclosure: this article was drafted with the assistance of Claude, the Anthropic product whose commercial moves are the central subject of the news pegs above. Over the four months it took to research and rewrite this piece, Anthropic restricted standard subscription access to third-party agents, dropped one of its team-tier products by a third, walked back the third-party restriction with a usage credit, and launched a small business product targeting the exact buyer this article is addressed to. Asked at each stage to argue against its own commercial interest in favour of open-source local-model adoption by Australian SMEs, Claude obliged without complaint and, if anything, with a little too much enthusiasm. Make of that what you will.