Strengths
- Z.ai reports a 50% gain over GLM-5.2 on Z.ai Code Bench, and at high effort it beat Claude Opus 4.8 there (31.4% vs 29.5%) using fewer output tokens.
- Z.ai reports 84.5% on the CyberGym security test, ahead of Mythos 5 (83.8%).
- Instead of a Chinese provider, you can use it through Mistral AI's API, which stores data in the EU by default.
- A one-million-token context and answers up to 128K tokens for $1.40 input and $4.40 output per million tokens.
Weaknesses
- It only understands text, so it cannot take images or scanned documents.
- Reasoning cannot be turned off, you can only set the effort to low, high or max.
- Z.ai itself admits the gap to the best closed models is wider on deeper vulnerability exploitation tests.
- The license is unclear: Hugging Face lists a custom glm-5.3 license, while Mistral's docs say MIT.
Best for
- agentic coding in Claude Code, Cline or OpenCode
- security code review
- an open model with EU data hosting via Mistral
Not ideal for
- images and scanned documents
- quick answers without thinking
How to use it in Europe
Through the Z.ai API in USD or the GLM Coding Plan subscription for coding tools. With Z.ai, check its terms for where your data is processed. For EU companies the Mistral AI API (model zai-glm-5-3) is more practical, as data is stored in the EU by default and input and output prices are the same. You can download the weights from Hugging Face and run them on your own server, but the model has around 753 billion parameters, so it needs powerful hardware.
What is coming
Rumor, unconfirmed Zhipu (Z.ai) founder Tang Jie has publicly mentioned a future model, GLM-6.0. He says self-improvement will be one of its directions. The company has not given a release date. When: no date, the A2Agent blog estimates early 2027 →
Price compared with similar models
Benchmark results
| Benchmark | Score | Source |
|---|---|---|
| DeepSWE v1.1 | 66.9 | docs.z.ai ↗ (vendor-reported), August 2026 |
| Terminal-Bench 3.0 | 28.3 | docs.z.ai ↗ (vendor-reported), August 2026 |
| CyberGym | 84.5% | docs.z.ai ↗ (vendor-reported), August 2026 |
| Surge AI blind human coding eval (cited by Mistral) | 3.60 / 5 | mistral.ai ↗, October 2026 |
Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.
More models from this vendor
- GLM-5.3-Flash: Z.ai's cheap, fast open model that understands text, images and video and, by Z.ai's account, comes close to top models in tests.
Similar models from other vendors
- Codestral 25.08 (Mistral AI): Mistral's model for code completion right in your editor. Fast and cheap, but dated for autonomous coding agents.
- Grok Build 0.1 (xAI): A fast Grok coding model and the cheapest Grok on the API, built for agents that write and fix code on their own.