We put GLM-5.2, the open-weight model from Z.AI, up against Claude Opus 4.8 on a real client project—same job, same tool, only the model changed. Here's what we found.
Last year we wrote about Claude Code with Opus, and the tool has since become part of how we build. More recently, much of the discussion online has been about GLM-5.2, an open-weight model that reportedly matches the best closed models at a fraction of the cost.
We work from Shanghai, so a capable model served from inside China interests us for reasons beyond price. We wanted to see whether the claims held up.
GLM-5.2 is an open-weight model from Z.AI, a Chinese AI company: the model itself is free to download, though running it takes serious hardware, so most teams—ourselves included—use it through Z.AI's own service. That service is much cheaper than the closed alternatives and runs from servers inside China—no VPN required.
Opus 4.8 is Anthropic's most advanced model and our usual default. We ran both inside Claude Code, the same terminal tool we covered last year, so the only variable between the two runs was the model itself.
The job was rebuilding a website for a furniture client. The old site dated back about eight years—React and Gatsby on the frontend, Contentful on the backend. Contentful is a hosted content system that had become expensive over time, and it didn't always perform well for visitors inside China.
We moved the frontend to Next.js and the content to Payload, which is free, open source, and can be hosted anywhere—including inside China. That was the real benefit for the client, and it's why this kind of migration makes up a large part of our work now.
We reset the code to the same starting point before each run, so both models began from the same place, and gave each one the same strict prompt: this is a technical migration only, follow the original design exactly, and reuse the existing frontend code unless there's a strong reason not to.
We also pointed both models at two of our past projects on the same stack and told them to follow the same structure and standards. We measured four things—whether it finished, whether it respected the design, how much work it used, and how fast it felt from our office.
Both models finished the migration. Both also made the same mistake at first: we told them not to redesign anything, and each changed a few things anyway.
One firm correction fixed it in both cases. Neither got it right on the first try, and neither needed a third.
Neither run was a one-shot. Both models got the site to roughly ninety percent on their own. The last stretch was detail work—checking each page against the original, correcting spacing and small layout differences, and connecting the smaller pieces of content by hand.
That part is normal. A migration is finished when every page matches the original, not when the first version runs, and the final stretch still takes an experienced person going through the site carefully.
The two results were close. Opus produced slightly cleaner work on one tricky part of the content model, and we made a small fix to the GLM version by hand.
Strip the labels and hand us the two finished sites, and we would struggle to tell which model wrote which.
We're on flat monthly plans for both, so neither run added cost on the day. But for anyone deciding which to pay for, price and speed are the real questions. Below is the work each model used, what it would cost at each company's published rates, and how long the working session took—treat the dollar figures as rough estimates.
Per use, GLM came out four to five times cheaper for this task, consistent with the claims that prompted the test. The caveat is that most teams, ourselves included, pay a flat monthly plan rather than per use—and a flat plan spread across all of your work narrows that gap.
This is the part that matters most to us, and it explains some of the time difference above. Reaching Opus from Shanghai means going through a VPN, and the connection is slow and sometimes drops. GLM runs from servers inside China—no VPN, no dropped connections—and it was simply faster to work with through the day.
For teams outside China this isn't a factor. For us, it's most of the reason we ran the test.
Claude Code is built by Anthropic and tuned for its own models, so this wasn't a perfectly neutral field. GLM is weaker at reading images, which matters for design work where you check the result by eye.
And one clean migration isn't a year of daily use. We'll keep testing it on real work before trusting it with everything.
For a defined, technical job like this migration, GLM-5.2 performed close enough to Opus 4.8 that the difference didn't change the result. It was much cheaper per use and, from where we sit, faster and more reliable to reach.