GLM-5.2 is Z.ai's (ZhipuAI's) flagship model for long-horizon agentic and coding tasks. It features a truly usable 1 million-token context window — roughly 5× larger than its predecessor GLM-5.1 — and supports up to 131,072 output tokens per response. Built on the GLM-5 Mixture-of-Experts backbone (744B total parameters, 40B active), it offers two thinking-effort levels (High and Max) for complex multi-step coding and engineering workflows. GLM-5.2 is compatible with eight agentic coding tools out of the box, including Claude Code, Cline, and OpenClaw, via an Anthropic-compatible endpoint.
Evaluations
14
across 12 benchmarks
Latency
26.6s
Context length
1M
tokens
Cost
$1.40 · $4.40
input · output per 1M tokens
Key takeaways
GLM 5.2 is Z.ai's flagship model for long-horizon agentic and coding tasks, featuring a 1 million-token context window and supporting up to 131,072 output tokens.
The model is built on a GLM-5 Mixture-of-Experts (MoE) backbone with 744B total parameters (40B active) and offers two thinking-effort levels (High and Max) for complex workflows.
It demonstrates strong mathematical problem-solving capabilities, achieving 100% accuracy in Counting & Probability, Number Theory, and Prealgebra on the MATH-500 dataset.
GLM 5.2's primary strengths include detailed, step-by-step solutions and consistent high performance across various mathematical topics, but it can struggle with precise algebraic simplification and adherence to strict formatting.
The model shows a 42.4% accuracy on the SWE-bench Pro (mini-swe-agent) dataset for multi-turn evaluations, indicating potential for improvement in complex coding agentic tasks.
