A Coding AI That Beats GPT-5.5 for One-Sixth the Price
A new open-weight AI from Beijing scores higher than GPT-5.5 on real coding tests and costs a fraction of the price — but teams should understand where their code goes before switching.
- 62.1
- GLM-5.2 score on SWE-bench Pro
- 58.6
- GPT-5.5 score on the same test
- ~1/6th
- cost vs GPT-5.5 on output tokens
- 753B
- parameters, MIT-licensed and free to download
A new AI model called GLM-5.2, built by Beijing-based Z.ai, scores 62.1 on SWE-bench Pro — a test where AI fixes real bugs in real software projects. That beats GPT-5.5's 58.6. It costs about one-sixth as much as GPT-5.5 to run. The model is also free to download under a permissive MIT licence, meaning any team can host and use it themselves.
The numbers: GLM-5.2 costs $1.40 per million input tokens and $4.40 per million output tokens via Z.ai's API — where tokens are the small chunks of text the model reads and writes. GPT-5.5 costs $5 and $30 per million. On third-party hosting platforms like DeepInfra, GLM-5.2 runs for under a dollar per million input tokens. For teams processing large volumes of code, this gap changes the maths entirely.
There is one real trade-off. Using GLM-5.2 through Z.ai's cloud means your code goes to servers in China, where local law can require that data be shared with the government on request. For sensitive projects, client code, or regulated industries, that matters. The cleaner path for those teams is to self-host: download the model weights and run it on your own servers under the MIT licence, at no cost.
Claude Opus 4.8 still leads on coding at 69.2 on SWE-bench Pro, so GLM-5.2 is not the top performer. But there is a difference between the best model and the best model you can afford to run at scale. Test GLM-5.2 on a low-sensitivity project this week — third-party hosts make it easy without any server setup. If it meets your quality bar, the savings follow.