GLM-4.6 Released: Zhipu Enhances Long-Context Coding and Search Agent
Zhipu released GLM-4.6 in September 2025, extending the context window from 128K to 200K, and enhancing coding, reasoning, and tool use.
Zhipu shifted its model iteration focus from parameter scale to real coding trajectories, context efficiency, and agent integration, starting to use public task trajectories to support engineering capability claims.
GLM-4.6 supports tool use during inference and a search agent; the company also publicly released CC-Bench trajectories and reported completing real coding tasks with fewer tokens compared to GLM-4.5.
Model vendors increasingly need to show verifiable task trajectories, costs, and failure distributions, rather than just final scores; coding agents are also becoming the main entry point for Chinese models to reach overseas developers.
Technical leaders should reuse the same set of internal repository tasks to compare different models, and record tokens, retries, modification accuracy, and manual review time.
Observe the coverage of public trajectories, success rate on complex repositories, long-context stability, and long-term retention in frameworks like Claude Code.