A Cheap Chinese AI Model Is Making Coders Rethink Their Bills
Z.ai's GLM 5.2 costs a fraction of Anthropic's top models and handles real work. Rate limits, hallucinations, and data-privacy questions mean it won't suit everyone.
Key points
- Z.ai released GLM 5.2 on 16 June 2025, a large language model with 753 billion total parameters but only 40 billion active at once, making it faster to respond.
- Z.ai charges $4.40 per million output tokens (the chunks of text the model generates), less than a fifth of the comparable price for Anthropic's Opus 4.8.
- GLM 5.2 is open-weights, meaning any organisation with its own hardware can download and run it without paying Z.ai at all.
- Stanford's AI Index found that Chinese labs produced just over half as many notable AI models in 2025 as US labs, up from roughly a third in 2023.
- In Z.ai's own benchmarks, GLM 5.2 completed just 13 percent of tasks in SWE-Marathon, a tough long-running coding test, while Anthropic's Claude Opus 4.8 doubled that score.
Zain Hasan is an AI engineer at Together AI and the kind of person who tracks exactly what he spends on AI tools. Hard problems go to a top-tier model, easy problems go to something cheaper. His cheaper pick right now is GLM 5.2.
GLM 5.2 comes from Z.ai, a Beijing-based AI lab. The model costs $4.40 per million output tokens via Z.ai's API (a programming interface that lets your app talk to their model). That's less than a fifth of the price of Anthropic's Claude Opus 4.8, and about a tenth of Anthropic's Fable coding model. First reported by IEEE Spectrum AI, the release landed on 16 June 2025.
It's also open-weights: Z.ai published the model's underlying data so anyone with enough computing hardware can download and run it themselves, for free.
Is GLM 5.2 actually good enough to trust?
For many tasks, yes, but with clear limits. David Nix, a principal software engineer at Denver-based MetaRouter, says GLM 5.2 comes close enough to top US models that it now handles 10 to 20 percent of his daily AI workload. Front-end design work is where he reaches for it first.
Hasan says the model holds a coherent train of thought across hours of back-and-forth conversation. Earlier cheaper models often lost the thread after just five to fifteen exchanges.
Not everyone had smooth results. Sai Kiran Myadaram, a software engineer at Bengaluru-based Indhic AI, burned through his weekly Z.ai token quota in under three days. He ran into hallucinations (where an AI confidently produces wrong or made-up answers) and excessive over-planning on simple fixes, and has since moved back to OpenAI's Codex.
The privacy question is real. GLM 5.2 comes from China, and some US companies worry about sensitive code passing through Chinese-linked servers. The open-weights option is the practical answer: host the model on your own hardware and your data never leaves your building. That's something most top US models can't offer; they sit behind an API with no self-hosting available. Our 14 July story on Grok Build quietly uploading codebases showed how little visibility engineers typically have over where their code goes.
Benchmarks paint a mixed picture. Z.ai's own research report shows GLM 5.2 trailing Anthropic's Opus 4.8 on most coding tests, though it nearly matches Opus 4.8 on some shorter agentic coding benchmarks (tests where AI software completes multi-step programming tasks on its own).
Should you worry about the cost of ignoring this?
The bigger issue, Hasan argues, is that most software teams don't know what their AI tools cost per project. When the company pays, engineers reach for the most powerful model by default. GLM 5.2 won't replace the best US models for the hardest jobs. But for routine work, the price gap is real and wide enough that ignoring it is starting to look like a choice, not an oversight.
