Last year everyone had ChatGPT open in a tab. This year it's Claude, but given the expected rise in token prices, I am starting to think next year's model will be Chinese open source.
But don't laugh. There's a new model in town called GLM-5.2 from Beijing. If you download it and run it yourself, it costs nothing. And it's not instant mash.
Z.ai, formerly Zhipu AI, was spun out of Tsinghua University in 2019 and listed in Hong Kong this January. Washington put the company on its trade blacklist a year ago, which tells you the US takes it seriously even if your board has never heard the name.
In June it released GLM-5.2, a 753-billion-parameter model with a one-million-token memory, under an MIT licence. That means you can feed it an entire codebase in one go and run a business on it without paying Z.ai a penny.
Only around 40 billion of those parameters fire for any given task, so it runs far cheaper than its size suggests.
I went to a breakfast last week with a senior Silicon Valley figure. One of his points was that the US is really concerned about Chinese technology because in many cases it is ahead, so much so that US companies are scrambling to hire Chinese developers and data experts. How the tables have turned.
In the UK, plenty of people still work on the assumption that the best AI will always sit behind a handful of US models paid for by the token. That AI is expensive, perhaps not for the end user yet but it will be, because it is spectacularly inefficient.
Ask a powerful model to fetch you a bag of crisps from the shop and it would start by researching the genome of the potato. It would then benchmark every frying method known to science, design the supply chain for a thousand boxes you never asked for, and eventually arrive at the shop having routed itself via Portugal. That's after writing 18 plans you had to correct.
GLM-5.2 could start to change this, because it is efficient and free to run.
The timing of its release was no accident. It landed in the same fortnight Washington banned foreign access to Anthropic's two most powerful models. It then caved, switching them back on this week in part because the ban was doing nothing except handing China a head start.
You cannot export-control something anyone can download for free, and China is taking full advantage.
Independent testers put these models through jobs that take a person days, fixing bugs across a whole codebase, or turning a scattered pile of emails and documents into a finished board deck. On the coding version GLM-5.2 beat OpenAI's best and came close behind Anthropic's Opus.
On the messy real-work version, run by a firm called Artificial Analysis, it finished third in the field, behind only Anthropic's two flagship models.
On that same test, Anthropic's flagship Fable 5 spent around $31 a task, its cheaper Opus model about $10, and GLM-5.2 did the same job for a little over $2, less than a tenth of what Fable 5 cost.
There is of course a catch - and I bet you've guessed it. Run Z.ai's model through its Chinese cloud and your data sits under laws that can compel the firm to hand it over, so the privacy prize only lands if you host it yourself.
And the same openness that frees the buyer frees the attacker, which is why security teams are watching these models closely. A self-hosted model still needs checking for what it might send to China before you trust it with anything sensitive.
Power in AI could soon slide from the companies guarding it to people who can own and run it themselves.
Whether it's Z.ai or something else, my bet is that open source is going to make huge progress by next year.
To get up to speed with AI in your company, whether it's learning the basics or where your governance model sits, check out Caipability - workshops to align and accelerate your team's AI knowledge.
