Cheaper AI models won't cut your costs on their own

A Beijing lab recently released an open-weight model that handles most everyday AI work about as well as the expensive frontier options, at roughly a fifth of the price, with the weights free to download and run. GLM 5.2 is the clearest example so far, but it sits inside a wave that includes DeepSeek V4 and a handful of others. For anyone paying a serious monthly bill for AI, the obvious move looks simple: switch to the cheaper model and pocket the difference.

We have spent enough time building and running these systems to know that the obvious move is exactly the part that goes wrong. The model is rarely the hard part. The hard part is everything wrapped around it, and that is where the cost and the real advantage sit.

Most of your AI work is more ordinary than it feels

When people picture AI, they picture the hard cases. In practice, the bulk of what businesses run through a model is ordinary, repeated work with a known shape: drafting a first pass of copy, summarising a long thread, pulling fields off an invoice, answering a support question that has been answered a thousand times before, generating standard boilerplate for a familiar coding task. These are tasks with lots of prior examples and outputs a human can check at a glance.

Cheaper open models are very good at this ordinary middle. On a lot of it they are close to the frontier, and on some design and front-end work they now lead outright. Where they still fall behind is the hard edge: long, multi-step problems where a small early mistake compounds over hundreds of steps, novel situations with no clear precedent, and the work where you want the most capable model you can get. The trouble is that almost no business has measured how its own workload splits between the ordinary middle and the hard edge. Until you have, you are guessing about which tasks can safely move to a cheaper model and which cannot.

The model is the cheap part

The instinct is to treat a model swap as a one-line change in configuration. In practice it means rebuilding the machine the model sits inside. A model on its own is a brain in a jar. It does nothing useful until you wrap it in a system: the prompts, the memory, the way tasks get broken down, the tool calls, the guardrails, and the retries for when something fails. That system gets tuned to the quirks of one model, and it does not lift and shift cleanly onto another.

Moving a real workload from one model to another is a project in itself. The evaluations that prove quality held have to be rebuilt. The same model has to be tested across different hosting providers, because the provider changes the results, and some of them quietly serve lower-quality, compressed versions of the weights. Prompts get re-tuned to recover the behaviour the old model gave you for free. Then the whole thing rolls out slowly while someone watches for a drop in quality. Done properly, it is weeks of engineering before the saving shows up.

That is what the cheaper sticker price hides. It covers the tokens and says nothing about the engineering around them. Cheap tokens are easy to buy. A cheaper system that still does the job is the part that takes skill.

Cheaper tokens are not the same as a cheaper bill

Getting value out of cheap models is an engineering problem before it is a procurement one. You need to know your task distribution. You need evaluations that tell you, for your work, where the cheaper model is good enough and where it quietly degrades. And for most setups you want routing: send the ordinary majority of tasks to the cheap model, and escalate the hard ones to a frontier model, automatically, on the fly. This is the same shape as the human-in-the-loop automation we build elsewhere, where most of the volume runs autonomously and the awkward cases get handed up for review.

None of that is trivial, and the people who can do it well are scarce. AI engineering talent is in heavy demand and priced accordingly, which is why many companies that would benefit from this work never get to it. The model got cheap. The skill to wire it into a reliable system did not. For an agency or an in-house team that can do that wiring, this is a real opening, because the difference between a cheap model and a cheap bill is exactly the part most businesses cannot build for themselves.

Where your data goes is part of the price

There is a catch underneath the catch, and it matters more in Australia than the headline price does. The cheapest way to use one of these open models is through the lab’s own hosted API. For the leading Chinese models, that API runs under Chinese jurisdiction, including laws that can compel a company to hand over data. For a business processing client records, contracts, or anything covered by the Privacy Act, that routing is not a detail.

The clean answer is to self-host the open weights on infrastructure you control. The practical answer is that the strong open models are enormous. Running GLM 5.2 at full quality needs on the order of 1.5 terabytes of GPU memory, roughly eight high-end data-centre cards working together. Most small and mid-sized businesses do not have that, and would not stand it up for a single workload. So there are really three options, and each comes with a cost: cheap tokens routed through someone else’s jurisdiction, more expensive tokens under a contract you trust, or a private deployment you fund and run yourself. Each is defensible. None is free.

The quiet trade of convenience for control

While the open models pull on price from below, the frontier vendors are pushing from a different direction, and it is the part of this story we would most want a business owner to think about.

The newest enterprise AI products do not sit in a separate window any more. They live inside the tools your team already uses. Anthropic recently put Claude directly into Slack as a shared teammate you tag like a colleague, one that follows your channels, remembers them over time, and works on tasks for hours on its own. Microsoft, Salesforce and Google are all building versions of the same idea. These products are very good, and that is the point. The more your team leans on one, the more of your company’s context, the half-stated decisions and the way work gets done, ends up living on a vendor’s infrastructure.

That convenience has a long tail. A model that has quietly absorbed months of your context becomes very hard to remove, no matter how cheap an alternative gets. Reporting on the launch, including in TechCrunch, framed the capture of organisational context, and the switching cost it creates, as part of the strategy. We use frontier models ourselves, and we are not warning anyone off them. The point is narrower: make the choice deliberately, because the default is to hand over your own operating knowledge, the most valuable thing you have, in exchange for a very convenient assistant.

What to do with all this if you run a business

The useful response is to get specific about your own situation. A few questions worth sitting down with before you change anything:

• Do you actually know how your AI workload splits between ordinary, repeatable tasks and the hard ones?
• For the ordinary majority, have you tested a cheaper model on your own work rather than on a public benchmark?
• Do you have access to the engineering skill to build the evaluation, routing and data handling that turns a cheap model into a cheap bill?
• Where is your company’s context accumulating, and are you comfortable with who owns the infrastructure it sits on?

Most businesses have not answered these, which is fair, because the ground moved fast. Cheap, capable AI is here, and the access picture for the very top models keeps tightening through export controls, which only makes the open options more relevant. The advantage now goes to whoever can take that cheap intelligence and build the reliable, private, well-routed system around it. That last part is the work, and it is the work we do.

If you are weighing up how to use AI without overspending or handing your data to the wrong place, that is exactly the problem our AI development and automation work is built around: secure, private systems that put the right model on the right task and keep your proprietary data yours.