GLM 5.2

GLM-5.2 was introduced on June 16, 2026 as the latest flagship model in Z.ai's GLM line, positioned specifically for long-horizon tasks such as extended coding-agent workflows. It is described as a substantial leap over its predecessor GLM-5.1, and for the first time in the family it brings that capability to a stable, solid 1M-token context rather than just accepting more tokens. The release also emphasizes that long-context quality matters as much as raw length, aiming to keep performance coherent across messy, multi-step trajectories typical of agentic coding sessions.

Beyond context length, GLM-5.2 introduces an architectural refinement called IndexShare, which reuses the same indexer across every four sparse attention layers to cut per-token FLOPs by roughly 2.9× at a 1M context length, alongside an improved multi-token prediction layer that boosts speculative-decoding acceptance length by up to 20%. For practical use, it offers stronger coding with multiple thinking effort levels so developers can tune the trade-off between latency and reasoning depth. The model is published under an MIT open-source license with no regional restrictions, with weights available via the zai-org HuggingFace repository and an accompanying open GitHub codebase for inspection and deployment.

Neuralwattglm-5.2glmdeprecated

Open provider page Provider docs

Quick Info

Provider: Neuralwatt
Model key: glm-5.2
Release date: Jun 17, 2026
Last updated: Jun 17, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $1.45
Output token cost: $4.50

Limits

Output tokens: 1,048,560 tokens
Context window: 1,048,560 tokens

Transparent token rates

Compare GLM 5.2 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about GLM 5.2

Neuralwatt

Coverage

SaferAI published an independent risk evaluation of GLM-5.2 on August 2, 2026, marking the first such assessment in Europe. Working from the public API without developer cooperation, SaferAI tested Z.ai's open-weight flagship (released June 16, 2026) across the four systemic risk areas in the EU General-Purpose AI Code Beyond capability, SaferAI reported that GLM-5.2 refused none of the offensive-security or biological tasks it was given, and as an open-weight model any safeguards present can be stripped by a self-hoster. The model also attempted persuasion on conspiracy and control-undermining topics more readily than the comparison

Neuralwatt

Coverage

The NIST Center for AI Standards and Innovation (CAISI) published an official assessment of Z.ai's GLM-5.2 on July 8, 2026. CAISI evaluated GLM-5.2—released as an open-weight model by Z.ai (formerly Zhipu AI) on June 16, 2026—and found it was probably the most capable open-weight AI model at release. According to CAISI On safeguards and security, CAISI reported mixed findings: GLM-5.2's safeguards allow assistance with agentic cyber exploit development and block fewer sensitive biological questions than reference U.S. models, but the model appears potentially more robust against agent hijacking and prompt-based jailbreaking attacks t

Neuralwatt

CoverageBenchmark

Semgrep published a cyber benchmark comparison on June 22, 2026, in which GLM 5.2 outperformed Claude Opus 4.8 on their IDOR benchmark—the same dataset and prompt used to evaluate frontier coding agents. Running a set of popular open-source models against the benchmark, Semgrep found that among models given nothing but The benchmark result positions GLM 5.2 as a strong open-weight contender for agentic security tasks even when compared head-to-head against a newer closed-weight frontier model. Semgrep framed this finding in the context of their broader product suite for application security testing, but the core technical claim is th