Six AI agents are now competing to run your business: Here's what each one actually costs

This piece is the practical follow-on. Six players are now active in this market. Five are large American technology companies. One is open source.

The choice between them, made consciously or by default, is one of the more consequential IT decisions an Australian small or medium business will make this calendar year.

The aim of what follows is to lay out, soberly and in Australian dollars, what each option offers, what it costs once GST and currency markup are included, and what each one quietly asks of you in return.

The six players

The agents currently positioning to run work across Australian businesses fall into three groups: the established commercial AI companies (Anthropic, OpenAI), the platform giants building consumer-grade competitors (Meta, Google) and the open-source alternatives (OpenClaw, plus an enterprise-hardened version called NemoClaw).

Anthropic/Claude Cowork: Anthropic is the maker of Claude, one of the two most capable AI models currently available to the public.

The company's agent product is called Claude Cowork. It is paid, polished,and intended for professional use.

Pricing in Australia, after currency conversion and the typical regional markup applied to US software, sits about A$30 per month for the entry tier and A$300 per month for the high-usage tier.

Anthropic recently restricted how its standard subscriptions can be used inside third-party agents, requiring users who want that flexibility to pay separately on a per-use basis through the company's API.

OpenAI/Codex and ChatGPT: OpenAI's equivalent products are Codex (for coding work) and ChatGPT itself, which now operates as an agent on the higher tiers.

Australian pricing is comparable - roughly A$30 per month for entry and around A$300 per month for the top consumer tier.

The company has explicitly endorsed using its subscriptions inside third-party agents like OpenClaw and others, and has hired the original founder of OpenClaw onto its team.

Of the commercial options, this is currently the most flexible posture toward customers who want to mix and match tools.

Meta/Hatch: Meta is reported to be building "Hatch,", an internal project described in leaked documents as a consumer version of OpenClaw.

It is targeted at non-technical users and is expected to be distributed through Meta's existing platforms (WhatsApp, Instagram, Facebook).

Internal testing is reportedly scheduled for next month. There is no public pricing yet, but the precedent across Meta's product line strongly suggests Hatch will launch free, supported by advertising or by upselling premium tiers later.

Google/Remy: Google is reported to be internally testing an agent codenamed Remy, powered by the company's Gemini AI.

Internal documents describe it as a "24/7 personal agent for work, school and daily life" that turns the existing Gemini app into something that can take actions on your behalf rather than just answer questions.

Industry sources have suggested Remy could be previewed at Google I/O.

Pricing is unconfirmed but a free tier integrated with existing Google Workspace subscriptions is the most likely launch model.

OpenClaw: The open-source agent we covered in our last piece is free to download, runs on hardware you own and works with whichever AI model you connect to it.

The cost to run it is the cost of the hardware (a Mac Mini, a small server, a sufficiently capable laptop) plus optional, capped charges for paid AI models if you want to use them for harder tasks.

A typical small-business setup runs at roughly A$20 to A$40 per month total for the AI compute, on top of the one-off hardware cost.

There is no subscription, no seat-based pricing and no vendor whose terms you depend on.

NVIDIA/NemoClaw: NemoClaw is an enterprise-friendly version of OpenClaw, with additional security and reliability features built by NVIDIA, the company that makes most of the world's AI hardware.

It is currently in alpha testing and free for early-access partners. Eventual pricing is expected to follow standard enterprise software conventions: meaningful annual contracts aimed at larger businesses rather
than SMEs.

What you actually pay over three years

Headline pricing tells you very little on its own. The number that matters for an SME considering any of these options is the total cost over a realistic ownership horizon.

Three years is a useful frame because it broadly matches both standard hardware depreciation and typical enterprise software contract lengths.

For a single user on the higher commercial tier (which most Australian SMEs running agentic work will require) three years of Claude Cowork or ChatGPT's premium tier comes to roughly A$10,800 per seat.

Add a second seat and you are at A$21,600. Add a team of five and the three-year commitment exceeds A$54,000.

None of that money builds equity. It rents access for the period it is paid, then stops.

For an OpenClaw setup running on a Mac Mini with moderate paid-AI usage on top, the three-year total for the same five-person team works out to about A$3,500 to A$5,500 (one Mac Mini at around A$1,200 to A$2,500 depending on configuration, plus around A$30 to A$50 per month in API charges across the team for the small share of work that benefits from frontier-tier AI).

The hardware retains some residual value at the end of the period and continues to function for as long as you want to keep using it.

The gap between those figures is roughly tenfold. That is the kind of difference that, in any other software category, would be the entire story.

In this category, it is not, because the commercial options offer convenience that the open-source path does not.

The question for any business considering this is whether that convenience is worth about A$50,000 over three years for a five-person team.

The honest answer depends on the team and the business; we will return to it next week.

The Australia tax, and the permanent write-off

Two financial considerations apply specifically to Australian businesses and tilt the maths above further.

The first is what the local tech industry calls the "Australia tax". Software priced in US dollars typically attracts not only GST and exchange-rate margin, but a regional markup that has historically run between 20 and 30 per cent above the equivalent US headline price.

This is the reason a product priced at US$200 in San Francisco frequently lands closer to A$340 here, even when a direct conversion at market rates would suggest A$310.

None of the commercial options on the list above are exempt from this dynamic. The open-source path, paying once for hardware bought through Australian retail channels, is.

The second is the instant asset write-off. In the 2026 Federal Budget to be handed down tonight, the Treasurer is expected to confirm that the $20,000 instant asset write-off for small businesses with aggregated turnover under $10 million is being made a permanent feature of the tax system.

This ends a decade of year-by-year extensions that have created planning uncertainty for small businesses since 2015.

Genuine business hardware purchases under $20,000, installed and ready for use, can now be deducted in full in the year of purchase rather than depreciated over several years. There is no longer an end-of-financial-year cliff to plan around.

The practical effect is that a A$2,000 hardware purchase is something close to a A$1,400 purchase after tax for a typical small-business buyer, every year, without a deadline.

The tax case for acting now is no longer about a window closing. It is about whether the hardware you need is actually on a shelf.

The Mac Mini supply situation 

A practical complication: the open-source path depends on having access to suitable hardware, and the most popular hardware for this purpose, the Apple Mac Mini, is currently in short supply.

On Apple's Q2 2026 earnings call, chief executive Tim Cook confirmed that the Mac Mini and Mac Studio "may take several months to reach supply demand balance".

Cook attributes the shortage to unexpectedly strong demand.

"Both of these are amazing platforms for AI and agentic tools, and the customer recognition of that is happening faster than what we had predicted," he says.

That is a meaningful disclosure. It is Apple's CEO confirming, on a publicly reported earnings call, that the local-AI thesis we are describing in this series is now mainstream enough to overwhelm Apple's own supply chain.

The base US$599 Mac Mini is currently listed as unavailable on the US Apple Store.

High-memory configurations of both the Mac Mini and Mac Studio, precisely the configurations most useful for local AI work, have been pulled from the store entirely, and lower-RAM units that remain listed are quoting four-month shipping delays.

EBay scalpers are listing units at A$1,100 to A$1,500 against a A$999 RRP.

The shortage is not a US-only phenomenon. In researching these articles, I have visited Apple flagship retail, three JB Hi-Fi locations and Harvey Norman in the past fortnight. None had stock in any meaningful configuration.

This presents a real problem for any business hoping to act on the open-source option.

The supply constraint is structural, driven both by AI demand for unified-memory machines and by a broader global DRAM shortage, and Apple's own CEO has signalled it will likely persist into the second half of 2026.

The window to act is no longer defined by a tax deadline. It is defined by whether the hardware you want is on a shelf at the moment you decide to buy.

Where the lock-in risks actually sit

A pricing comparison is incomplete without a sober look at what each option asks of you in return for the convenience it offers.

Three of these are worth understanding before any commitment.

Data: Each commercial agent learns about your business by accumulating context from the work it does for you. Over time, that context becomes valuable.

A year of email patterns, accounting workflows, client correspondence and document drafts is not trivial to recreate.

If the agent storing that context is a commercial product, that data lives on the vendor's servers and is governed by the vendor's terms.

Migrating it to a different vendor, or pulling it back into your own systems, ranges from awkward to practically impossible.

Workflows: Once a team has built routines around a specific agent (the prompts, the integrations, the small habits of how staff hand work to the AI), switching costs rise sharply. This is not a unique feature of AI software, but the rate of habit formation in agentic workflows is unusually fast.

A six-month-old workflow involves muscle memory across an entire team that does not transfer cleanly.

Pricing terms: AI is an unusually young commercial category. None of these companies have stable long-term pricing models yet.

Anthropic dropped one of its team-tier products by one third earlier this year with no negotiation; the same dynamic could move pricing the other way at any of the commercial vendors with similar abruptness.

A subscription that looks affordable in May 2026 may look very different in May 2027, and the customer signing up today has no leverage over what changes between now and then.

The open-source path does not eliminate any of these risks entirely (your data still lives somewhere, your workflows still build over time) but it does keep the decisions about all three within your own four walls.

A note on what is not in this comparison

This piece has not assessed the relative quality of the AI doing the work inside each agent.

That is an unstable comparison; the AI models are improving every few months, and benchmarks that were authoritative six months ago are out of date now.

For most Australian SME workloads (drafting, summarising, reconciling, scheduling, classifying) all of the agents on this list will perform competently.

The differentiator for SMEs is not which one is smartest. It is which one fits the business's economics, data posture and risk appetite.

What we'll cover next

This article has laid out the field. In our final piece in this thread we will move from comparison to opinion: which of these six options we think makes the most sense for Australian SMEs, why the next quarter may matter more than the next four years, and what the playbook looks like in practice for a typical 20-person business.

If you have made it this far, you are now better-equipped to evaluate the AI coverage that will dominate the business pages over the next month than most of the people writing it.

That is, in itself, a useful place to be.

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.

A footnote, in the spirit of full disclosure: this article was drafted with the assistance of Claude, the Anthropic AI product mentioned above.