An open-weights AI model just beat GPT-5.5 on key coding benchmarks, and it costs roughly a sixth of what OpenAI charges. Z.AI’s GLM-5.2, unveiled between June 13 and 16, is designed for the kind of long, complex coding tasks that separate toy demos from production-ready tools.
The model edges out GPT-5.5 by approximately 1% on FrontierSWE and ranks first among all open-source models on long-horizon coding benchmarks. It also posts strong numbers on PostTrainBench and SWE-Marathon, two benchmarks that test a model’s ability to handle multi-step engineering problems that unfold over extended interactions.
What GLM-5.2 actually brings to the table
GLM-5.2 uses a Mixture-of-Experts (MoE) architecture. Instead of running all 744 to 753 billion parameters on every query, it activates roughly 40 billion at a time, routing each task to the specialists best equipped to handle it.
The context window jumped to 1 million tokens from the 200,000 tokens available in the predecessor GLM-5.1. For reference, 1 million tokens is roughly the equivalent of feeding the model an entire large codebase and asking it to reason about the whole thing at once.
On standard benchmarks, GLM-5.2 scores 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro. The model also supports two effort levels, labeled “High” and “Max,” letting developers trade off between speed and performance depending on the task.
The full weights are already live on Hugging Face under the handle zai-org/GLM-5.2, released with an MIT open-source license, meaning developers can use, modify, and commercially deploy the model with essentially zero restrictions.
Z.AI’s rapid iteration strategy
Z.AI, the company formerly known as Zhipu AI, has been on a tear in early 2026. GLM-5.2 follows GLM-5 and GLM-5.1 in a rapid-fire release cadence. API access and full open weights launched shortly after the initial reveal, alongside specific pricing tiers for developers who prefer hosted access over running the model themselves.