GLM 5.2: AI at a fifth of the cost, without selling out quality or your data

An open-weight model that holds its own against top-tier models on many tasks, at a fraction of the price. This isn't insider news: it's the cost of getting a task done with AI collapsing, and it changes the math of every automation.

For the first time a small business can move most of its AI work, research, drafts, first versions of interfaces, high-volume automations, onto a model that costs about a fifth, keeping the top-tier one only where it really has to reason. The bill drops, the quality on the right tasks doesn't. The model is called GLM 5.2, and it's open-weight.

The before: one premium model for everything

Many companies today pay for a single high-end model for everything, including research, first drafts and automations where the top tier isn't needed. Two consequences: an inflated bill and dependence on one vendor, who can raise prices or remove a feature whenever they want. And it can vanish for reasons outside its control too: a few weeks ago a top-tier model was switched off overnight on a directive. What you rent can stop being there, with no mistake on your side.

What GLM 5.2 is, in short

It's an open-weight model released in June under an MIT license: you can use it, host it wherever you want, and keep your data in-house. Under the hood it's large (a mixture-of-experts architecture, 753 billion total parameters, context up to a million tokens), but the number that matters for a business owner is another: on many practical tasks, from frontend to knowledge work to automations, it holds up against the closed top-tier models.

The part that changes the bill: price

The numbers speak for themselves. GLM 5.2 costs roughly 1.40 and 4.40 dollars per million tokens, input and output. A top-tier model like Opus 4.8 costs 5 and 25. On output that's about a fifth of the cost, on input even less. How much you actually save depends on your workload's mix, but the order of magnitude is that: a fraction. If half the work you pay premium rates for today can run here with the same output, your automation margins flip.

The honest limit

It doesn't win everywhere, and anyone selling it as "replace everything" is selling you smoke. On pure reasoning, complex debugging and the subtlest edge cases a model like Opus stays more precise. That's why the right answer isn't "switch to the cheap model", it's "use each model where it pays off".

How you actually do it

You route each task to the right model. The high-volume, low-criticality part goes to a cheap model like GLM, the steps that need judgment stay on Opus. For a developer it's a technical detail (you change one setting and the model runs inside the same environment, with nothing to rebuild); for the business owner it's just the bill going down. That's how Aima, the system I run for myself and for clients, is built: model-agnostic, the model is a choice you can change, not a cage. And with an open-weight model there's a bonus an owner feels: no lock-in, and the data can stay yours.

I'm already testing GLM inside my agents, on the steps where the quality is enough and the volume is high, keeping Opus where reasoning is needed. If you want to figure out where, in your company, you can cut the cost of AI without losing output, in a consultation we map your tasks, define the routing rule and estimate the saving on your current spend.