China's Z.ai GLM-5.2 tops OpenAI’s GPT 5.5 model on key benchmarks

Chinese startup Z.ai, formerly known as Zhipu AI, has unveiled GLM-5.2, a large language model (LLM) designed for "long-horizon" autonomous coding and engineering tasks.

Long-horizon tasks are projects that require an AI system to plan ahead, retain context, and complete many connected steps over an extended period, rather than carrying out a single action.

According to Ollama, which hosts Z.ai's flagship AI models, including the GLM-5 series, the model is built to handle large-scale software development work.


"With a truly usable 1M-token context window, it can handle project-level engineering context, execute long-running tasks more reliably, follow engineering standards more consistently, and complete the full development workflow, from requirements to multi-platform deployment, in a single task," said Ollama in a blog post.

Large context window, open access

A key feature of GLM-5.2 is its stable 1-million-token context window. A token context window refers to the amount of text a model can process and remember at one time. A larger window allows the system to work with lengthy documents, large codebases, and complex tasks without losing context.

The company is also offering enterprise subscriptions starting at $12.60 per month.

Z.ai has released the model under an unrestricted MIT open-source licence, allowing enterprises to freely download it from open-source AI platforms such as Hugging Face.

An MIT open-source licence means there are no regional restrictions and users can access and modify the model without technical barriers.

Benchmark performance

The launch comes at a time when some enterprises are seeking alternatives to leading American proprietary AI models. Interest in open-source options has increased following the Trump administration's export-control directive last week that barred foreign nationals from using Anthropic's new Claude Fable 5 model.

On industry benchmark tests, GLM-5.2 outperformed most leading open-source models, including DeepSeek v4, and scored close to or above closed-source rivals such as OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8.

For instance, on Humanity's Last Exam (with tools), GLM-5.2 scored 54.7, ahead of GPT-5.5's 52.2 and close to Claude Opus 4.8's 57.9.

Humanity's Last Exam (HLE) is a benchmark designed to measure an LLM's reasoning abilities rather than simple pattern recognition. It consists of thousands of expert-level questions covering mathematics, science, humanities, and reasoning.

Another benchmark, FrontierSWE, evaluates whether an AI agent can complete open-ended technical projects lasting from several hours to tens of hours, including systems optimisation, large-scale coding, and applied machine learning research. On this test, GLM-5.2 trailed Opus 4.8 by just 1%, while outperforming GPT-5.5 by 1% and Opus 4.7 by 11%.

Developer-focussed plans

To help developers use the model in production, Z.ai has launched coding plans:

· Lite: $12.60 per month ($151.20 annually from the second year), aimed at smaller repositories.

· Pro: $50.40 per month, offering five times the usage allowance of Lite for day-to-day development on mid-sized repositories.

· Max: $112 per month, offering 20 times Lite's usage allowance along with dedicated resources during peak periods.

This comes as Zhipu AI (Knowledge Atlas Technology JSC), said earlier this month that it plans to seek a domestic listing on the Shanghai Stock Exchange's Sci-Tech Innovation Board.