OpenAI now gives away two capable models. GPT-OSS-120B and GPT-OSS-20B can be downloaded, fine-tuned and run wherever you like: your own servers, a private cloud, or a laptop in the case of the smaller one. If your business has only ever bought AI as a monthly subscription, that is a new kind of decision, and the word "free" is doing a lot of work in the headlines.
This is a note for owners and operations leads, written as at October 2026. Our short view: GPT-OSS deserves a place on your list of options, and for most Australian small and mid-market teams it does not deserve a migration project this quarter.
Is GPT-OSS actually free for an Australian business?
GPT-OSS is free to download and its licence permits commercial use without royalties, but it is not free to run. An Australian business that self-hosts GPT-OSS-120B or GPT-OSS-20B pays for GPU hosting, security patching and monitoring every month, and takes on its own privacy safeguards under the Privacy Act, because no vendor support, uptime guarantee or help desk comes with the weights.
Both models are built for reasoning, coding and tool use. The 120B model targets server-grade GPUs. The 20B model is small enough for a single high-end workstation, which we covered in our GPT-OSS-20B single-GPU deployment guide. What you get is a file of weights and a licence. Everything a subscription normally hides is now on your side of the fence.
What you take on when nobody is selling it to you
With managed Claude, a large share of the price pays for things you never see. Take the vendor away and those jobs do not disappear. They land on a named person in your business, or on a contractor's invoice.
Infrastructure. Someone owns patching, scaling and monitoring, including the week they are on leave.
Data handling. There are no enterprise data terms to point to. You write and audit your own safeguards.
Licence reading. Permissive is not the same as unconditional, so read the terms before a customer-facing product ships.
Quality. Benchmark scores from OpenAI's own testing are a starting point, not a promise about your documents and your customers.
None of this is exotic work. It is ordinary IT operations. The catch is that it is a fixed cost: you pay it whether the model answers one question a day or a million.
The spend level where the maths starts to work
A Sydney firm paying $1,200 a month for a managed AI subscription is not the buyer GPT-OSS was built for. The economics only begin to work once monthly API or subscription spend clears roughly $8,000 to $10,000 and stays there. Below that line, the fixed cost of hosting swallows whatever you save on usage fees.
| Monthly AI spend | What it suggests | Sensible next step |
|---|---|---|
| About $1,200 | Fixed hosting costs would exceed the whole bill | Stay on managed Claude |
| About $4,000 | Still well under the line | Stay managed and start tracking spend by workload |
| $8,000 to $10,000, sustained | Self-hosting may begin to pay | Run a costed comparison on one workload |
| Above $10,000 for three to six months | A real second option | Pilot GPT-OSS beside Claude, not instead of it |
The spend bands at $1,200 and $8,000 to $10,000 are the planning figures we use. The $4,000 row is illustrative. Your own line depends on GPU pricing, which moves, as we set out in our piece on GPU spot prices and the self-hosting break-even. You can put your own numbers through the ROI calculator before anyone books a workshop.
Why "stays there" matters
One big month is not a trend. A tender response, a document backlog or an end-of-financial-year rush can push usage up and let it fall straight back. Hosting costs do not fall back with it. That is why three to six months of usage data is the right evidence, and a single invoice is not.
Four questions to settle before you commit
Before a self-hosted model touches real work, we would want a written answer to each of these.
What does monthly AI spend look like across every tool in the business, not only the headline subscription?
Who, by name or by vendor, owns security patching for the self-hosted model?
What happens to customer data if the instance fails during business hours in Melbourne or Brisbane?
Does your industry carry obligations, such as APRA CPS 234 in finance, that a self-hosted setup makes harder to meet?
If the second question has no owner, stop there. An unpatched model server with access to customer records is a worse position than the subscription you started with.
What not to conclude from the release
Two readings of GPT-OSS are common and both go too far. The first is that subscriptions are now a waste of money. They are not: at small and mid-market volumes a managed Claude deployment, with support, guardrails and data handling commitments already in place, is usually the cheaper total bill. The second is that open weights are a toy. They are not that either. A second supplier you can run yourself is real bargaining power and a real fallback.
The middle path is the one we see working. Keep Claude for the work where judgement and accountability matter, and test an open-weight model on one narrow, high-volume task once the spend justifies it. We described that pattern in running Claude and an open-weight model side by side.
Open-weight releases like GPT-OSS are a milestone. They are not, on their own, a reason for a 15-person Australian business to rebuild its AI stack this quarter. If you would like a second pair of eyes on your spend and where the line sits for you, book a short call with us.
