Z.ai Launches GLM-5.2 with 1M-Token Context Window and New Thinking Effort Levels
Z.ai has released GLM-5.2, its latest large language model, on June 13, 2026, making it available across all GLM Coding Plan tiers. A key feature is its 1-million-token context window, which significantly expands the model's working memory for coding agents. This new iteration also introduces 'High' and 'Max' thinking-effort levels, with 'Max' recommended for complex, multi-step coding work. GLM-5.2 is designed to integrate with platforms like Claude Code, Cline, and OpenClaw via an Anthropic-compatible endpoint. While Z.ai did not provide performance benchmarks at launch, the company has stated that MIT open weights for the model are expected to be released the following week.
Z.ai officially launched GLM-5.2 on June 13, 2026, making the large language model available across all GLM Coding Plan tiers. This marks the third major release in the GLM-5 series, following GLM-5 (February 11), GLM-5-Turbo (March 15), and GLM-5.1 (April 7).
The standout feature of GLM-5.2 is its 1,000,000-token context window, identified as glm-5.2[1m] in its configuration. This represents a substantial fivefold increase from GLM-5.1's approximately 200,000-token window. The expanded context allows a coding agent to hold an entire mid-sized repository in working memory, including source files, tests, configuration, and conversation history, thereby reducing the need for constant summarization. Each response from GLM-5.2 can generate up to 131,072 output tokens.
GLM-5.2 also introduces two thinking-effort levels: High and Max. Z.ai recommends the Max effort level for complex, multi-step coding tasks. In Claude Code, this setting is controlled by the `/effort` command, with options such as xhigh, max, and ultracode mapping to GLM-5.2's Max effort.
While Z.ai did not specify GLM-5.2's architecture in its launch materials, community notes suggest that the GLM-5 base model is a 744-billion-parameter Mixture-of-Experts (MoE) model that activates 40 billion parameters per token. GLM-5.1 maintained this architectural backbone with retargeted post-training.
Z.ai did not publish any benchmark scores for GLM-5.2 at launch, including SWE-bench, Terminal-Bench, or Code Arena numbers. The announcement focused on the model's availability, context capabilities, and an open-source roadmap. The company has committed to releasing MIT open weights for GLM-5.2 next week.
Potential applications for GLM-5.2 include whole-repository refactors, where the agent can manage cross-file dependencies within a single context window. It is also designed for long-horizon agent runs, aiming for sustained plan, execute, test, and fix loops. GLM-5.1, its predecessor, had sustained approximately 1,700 agent steps in one session and ran autonomous loops for up to eight hours. Additionally, GLM-5.2 can serve as a drop-in replacement for Claude Code and is capable of analyzing large documents that exceed 200,000 tokens.
According to Marktechpost, GLM-5.2 supports 8 agentic tools from day one.
