Z.ai released GLM-5.2 on June 13, 2026, the fourth flagship model in the GLM-5 line since February. The headline specification is a 1-million-token context window, a 5x increase over GLM-5.1’s roughly 200,000-token ceiling. Each response can return up to 131,072 output tokens, according to MarkTechPost.

What Shipped

GLM-5.2 is available across all GLM Coding Plan tiers: Lite, Pro, Max, and Team. The model exposes an Anthropic-compatible API endpoint, meaning agentic coding tools that already support Anthropic’s format can integrate by changing the base URL and model identifier alone. Z.ai lists eight tools with day-one compatibility: OpenClaw, Claude Code, Cline, OpenCode, and four others.

The release adds two configurable thinking-effort levels, High and Max, where Z.ai recommends Max effort for complex, multi-step coding work.

Release Cadence

The GLM-5 line has moved fast. GLM-5 launched on February 11, GLM-5-Turbo on March 15, GLM-5.1 on April 7, and GLM-5.2 on June 13. That is four flagship-tier coding releases in roughly four months. Z.ai has also announced that GLM-5.2’s model weights will ship under an MIT license, though no specific release date has been confirmed beyond “next week,” per the MarkTechPost report.

The Benchmark Gap

Z.ai published no benchmark scores at launch. There are no SWE-bench, Terminal-Bench, or Code Arena numbers. For comparison, GLM-5.1 scored 58.4 on SWE-bench Pro when it launched in April. The absence of benchmarks at release is unusual for a coding model competing against Claude, GPT, and Gemini variants that all ship with evaluation data. Z.ai focused the announcement on availability, context specifications, and the open-weight roadmap rather than performance claims.

Context Window in Practice

A 1M-token window changes what a coding agent can hold in working memory during a single session. An entire mid-sized repository, including source files, tests, configuration, and conversation history, fits without the forced summarization that smaller context windows require. According to MarkTechPost, GLM-5.1 sustained roughly 1,700 agent steps in a single session, running autonomous loops for up to eight hours. Z.ai positions GLM-5.2 as extending that trajectory, though its own endurance numbers remain unpublished.

The Drop-In Factor

The practical value for agent builders is the Anthropic-compatible endpoint. Teams running OpenClaw or Claude Code workflows can test GLM-5.2 by swapping the base URL in their configuration. No harness changes, no SDK updates. That lowers the evaluation cost to near-zero for teams that want to benchmark the model against their current provider in production-like conditions. It also means that if API access to a frontier provider gets disrupted, whether by pricing changes, rate limits, or export restrictions, GLM-5.2 becomes a swap-in fallback without retooling.

The GLM-5 architecture is reportedly a 744-billion-parameter Mixture-of-Experts model activating 40 billion parameters per token, based on community documentation cited by MarkTechPost. Z.ai has not confirmed whether GLM-5.2 modifies this base architecture or retrains on the same backbone with updated post-training.