Breaking

China's Zhipu Releases GLM-5.2, Rated on Par With Anthropic in Cyber Tasks

July 8, 2026 at 12:07 EDT

Foundation Models
Open Source
Security

GLM-5.2, an open-weight model released by China's Zhipu AI (Z.ai) in mid-June 2026, has drawn a series of assessments finding it matches Anthropic's frontier model in cybersecurity, rapidly narrowing the US-China gap in frontier AI. Because it was shipped as an open-weight model that anyone can download, modify and run locally, it is putting pressure on US national security and export-control policy.

June 2026 · Zhipu AI (Z.ai)

China's Open-Weight GLM-5.2 Pulls Even With Frontier US Models in Cyber

A freely downloadable MoE model matches Anthropic's frontier system on key security benchmarks — at a fraction of the cost — narrowing the US–China AI gap and rattling export-control policy.

744–753B

MoE parameters, MIT license, ~1M-token context

~$0.17

Cost per vulnerability detection

~1/6

API price vs US frontier models

Head-to-head: GLM-5.2 vs US frontier

Higher is better — GLM-5.2 leads on both.

IDOR detection — F1 score

39%

GLM-5.2

32%

Claude Code

SWE-bench Pro — coding

62.1

GLM-5.2

58.6

GPT-5.5

The story in one line

Open weights + frontier-grade cyber performance + rock-bottom cost = a governance problem with no easy off-switch.

Downloadable

Anyone can run it locally — no provider-side logging or abuse blocks.

Built under controls

Reached parity despite US chip export restrictions.

Momentum

Builds on DeepSeek, Qwen and Moonshot.

Why developers back it

Strong performance at very low cost
Repository-scale code analysis
Agentic, autonomous task workflows
Called the first open-weight model worth recommending for frontier cyber

Why officials are alarmed

Public weights defeat abuse-prevention controls
Reported misuse attempts
US warned of losing its cyber edge
Still trails US models on general tasks + dedicated pipelines

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.