For two years the advice was simple: if you want frontier quality, use a managed API. In 2026 that advice needs an asterisk. Open-weight models such as GLM-5.2, MiniMax M3, and Google's Gemma 4 now sit close enough to the top closed models that the old line, open is always behind, no longer holds. For an Australian business that quietly changes the maths on a few specific workloads. It does not change the answer for most of them. Here is a balanced read for teams weighing the two paths without the hype.
What actually changed
The shift is not that open models got a little better. It is that the gap at the top narrowed to the point where, on many everyday tasks, you would struggle to tell the output apart. A free, downloadable model that scores near the frontier removes the one objection that used to end the conversation early. What it does not remove is everything that sits around the model in production, and that is where the real cost and risk still live.
The case for self-hosting
Open weights have real, concrete advantages when the conditions are right:
Data residency you can prove. A model running on Azure Australia East or your own rack never sends a token offshore, which matters under the Privacy Act and for some government tenders that ask you to attest to it.
No per-token bill. Once the hardware is paid for, extra usage is close to free, so very high volume gets cheaper per request.
Full control. You can fine-tune on your own data and freeze a version so the behaviour never changes underneath you mid-quarter.
For a Brisbane logistics firm running millions of routine classifications a day, those points can add up to real savings, and the case deserves a serious look.
The case against
The download is free. The deployment is not. The honest costs include:
A production server. Australian self-hosted setups run roughly $1,500 to $3,000 a month for a small single-GPU box, and $15,000 to $40,000 a month for a high-availability cluster.
Engineering time. Expect two to four weeks of senior setup and 10 to 20 hours a month of upkeep, every month, not just once.
You own the safety problem. There is no vendor to catch a harmful output, patch a jailbreak, or answer a regulator for you.
None of these show up in the headline comparison of a free model against a paid one, which is exactly why so many self-hosting business cases look better on a slide than they do after a year in production.
How to decide
The deciding factor is rarely the benchmark. It is your token volume and your data rules. A useful rule of thumb from 2026 deployments:
Under about 5 million tokens a day, a managed API like Claude almost always costs less all in.
Between 10 and 30 million tokens a day, the maths gets close and depends on your setup and how much idle GPU time you carry.
Above 100 million tokens a day, self-hosting can cut 60 to 70 per cent off the bill.
Most Australian SMBs sit well under the first threshold, which is why we still reach for a managed model by default and treat self-hosting as a deliberate, evidence-led choice rather than a reflex.
The practical path
You do not have to pick one side forever. Start on a managed model, measure your real token volume for a quarter, then revisit with numbers in hand. If a single high-volume task is driving most of your spend, that task, and only that task, may be worth moving to an open model while everything else stays put.
The mistake we see most often is treating this as an identity choice rather than a cost one. Open versus managed is not a statement about how serious your engineering team is. It is a line on a spreadsheet that moves with your volume. Re-run the line every six months and let it decide for you.
The switching-cost trap
One number rarely makes it onto the business case: the cost of changing your mind. A team that jumps to a self-hosted model and later finds the upkeep too heavy faces a second migration to undo the first. Each move costs weeks of engineering and carries its own risk of breaking something that already worked. Counting that round trip changes how eager you should be to switch on the strength of a single leaderboard result, because the leaderboard never pays for the rebuild that follows it.
There is a quieter organisational cost too. Every model you run is a thing your team has to understand, monitor, and explain to an auditor or a nervous client. Two models is more than twice the surface area of one, because now you also own the routing logic and the question of which system did what when an answer comes out wrong. Simplicity has a dollar value that rarely shows up on a slide until the week it saves you a very bad day.
So the honest default for most Australian SMBs is to start on a managed model and stay there until a specific, measured workload makes the move worth it. Treat self-hosting as a decision you earn with data, not one you reach for because open weights are free and the demo looked impressive on a Friday afternoon. Measure your real volume first, switch once, and only where the numbers are loud enough to drown out the excitement.
Not sure where your workload sits on that curve? Book a free brainstorm and we will map it with you.
